6529 Mobile needs iOS 16.4 or later

Update in Settings > General > Software Update, then open 6529 Mobile again.

6529 needs a newer browser

Update your browser to its latest version, then open 6529 again.

Updating to the latest version

Updating to the latest version

Mise à jour vers la dernière version

Actualizando a la última versión

Aktualisierung auf die neueste Version

Test release0.1.0-rc.2

Every model

227 open-weights models. 139 run in the app today.

Untested test release. Expect problems, and don't rely on it for anything that matters. Known issues

How to read this page
  • Runs in the app today: its engine is llama.cpp, whisper.cpp or stable-diffusion.cpp, and the app serves what it makes: chat, embeddings, images or speech to text. The other models wait for engines and routes the app doesn't have yet.
  • Memory is the app's own estimate: the files' size divided by 0.85. A model fits when its files take at most 85% of the GPU's memory, or of the memory a Mac shares with its GPU, and the rest holds the context. A longer context needs more, and the engine has the last word. Without a GPU, the app suggests chat models of up to 5.5 GB, and they run slowly.
  • Scores are the publisher's own, from its model card at the pinned version, each linked. A percent is rounded to one decimal; a score out of a fixed maximum says so ("8.5 /10"); any other is shown as its card gives it. † marks the original model's score: a quantized or changed version can score differently. Cards use different settings, so don't rank models by them.
  • Measured by 6529 AI marks the runs 6529 timed itself, with the machine and the day.
  • Newest first goes by the day each model's repository was made on Hugging Face.

Filters

Not sure what fits? The app's Start page tells you.

227 of 227 models

† the original model's score

General chat, small

19 models, 18 in the app today

Assistants under 10 billion parameters. By the app's estimate, the smallest version of each needs 0.5 to 12 GB of GPU memory.

  • SmolLM2-360M-Instruct

    Hugging Face

    A small English chat model from Hugging Face with 360 million parameters, trained on 4 trillion tokens.

    Runs in the app today · Chat

    Parameters
    360M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.4 GBneeds 0.5 GBRuns in the app today

    Scores

  • Qwen3-0.6B

    Qwen

    A small general chat model from Qwen's Qwen3 family with 0.6 billion parameters.

    Runs in the app today · Chat

    Parameters
    600M
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today

    Measured by 6529 AI

    • Ran a test chat. · Intel Xeon, 16 vCPUs, no GPU · Nov 14, 2026
  • Qwen3.5-0.8B

    Qwen · packaged by ggml-org

    A small general chat model from Qwen's Qwen3.5 family with 0.8 billion parameters, for questions, writing help and simple tasks.

    Runs in the app today · Chat

    Parameters
    800M
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 1.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.8 GBneeds 1.0 GBRuns in the app today

    Scores

    • MMLU-Pro29.7% (the original model's score)
    • MMLU-Redux48.5% (the original model's score)
    • IFEval52.1% (the original model's score)
    • MMMLU34.1% (the original model's score)
  • OLMo 2 1B Instruct

    Ai2

    A 1-billion-parameter English chat model from Ai2, released with its code, checkpoints, logs and training details.

    Runs in the app today · Chat

    Parameters
    1B
    Context
    4,096 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 1.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.6 GBneeds 1.9 GBRuns in the app today

    Scores

    • GSM8K68.3% (the original model's score)
    • IFEval70.1% (the original model's score)
    • MMLU40.0% (the original model's score)
    • MATH20.7% (the original model's score)
    • BBH35.0% (the original model's score)
  • LFM3-1.7B

    Liquid AI

    A 1.7-billion-parameter text model from Liquid AI for local agents, tool calling, structured output and fine-tuning to one domain.

    Runs in the app today · Chat

    Parameters
    1.7B
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.2 GB
    License
    LFM Open License v1.0 (conditional)
    License terms

    Every copy carries the license and its notices. Commercial use is licensed only to an entity with less than 10 million US dollars of annual revenue.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.9 GBneeds 2.2 GBRuns in the app today

    Scores

  • Ministral 3 3B Instruct

    Mistral AI

    Ministral 3 3B Instruct is a 3-billion-parameter chat model from Mistral AI for small devices that also reads images.

    Runs in the app today · Chat

    Parameters
    3B
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF2.2 GBneeds 2.6 GBRuns in the app today

    Scores

  • Qwen3-4B (MLX, 4-bit)

    Qwen

    A general chat model from Qwen's Qwen3 family with 4 billion parameters.

    Not in the app yet · Its engine, MLX, is not built: the app refuses to start an MLX model.

    MLX is not built: the app refuses to start a model whose engine is MLX. The same model runs today as qwen3-4b-gguf.

    Parameters
    4B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    4-bit safetensors2.2 GBneeds 2.6 GBNot yet
  • Qwen3-4B

    Qwen

    A general chat model from Qwen's Qwen3 family with 4 billion parameters.

    Runs in the app today · Chat

    Parameters
    4B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 3.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF2.5 GBneeds 3.0 GBRuns in the app today

    Measured by 6529 AI

    • 56.2 tokens a second · Intel Xeon, 96 vCPUs, no GPU · Nov 14, 2026256 tokens at temperature 0 after a warm-up.
    • 33 tokens a second · Apple M2, 24 GB · Nov 14, 2026
    • Ran a test chat, a stream, a tool call and embeddings. · Graviton4, 16 vCPUs, no GPU · Nov 14, 2026
  • MiniCPM5-2B

    OpenBMB

    A chat model from OpenBMB with about 2.5 billion parameters, built for local assistants, coding agents, tool use and reasoning.

    Runs in the app today · Chat

    Parameters
    2.5B
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 3.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.7 GBneeds 3.2 GBRuns in the app today

    Scores

  • SmolLM3-3B

    Hugging Face · packaged by ggml-org

    A 3-billion-parameter chat model from Hugging Face that answers directly or thinks step by step first.

    Runs in the app today · Chat

    Parameters
    3B
    Context
    65,536 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 3.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF3.3 GBneeds 3.9 GBRuns in the app today

    Scores

  • Phi-4-mini-instruct

    Microsoft · packaged by unsloth

    A 3.8-billion-parameter chat model from Microsoft, trained on filtered web and synthetic data with a focus on reasoning and math.

    Runs in the app today · Chat

    Parameters
    3.8B
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 4.9 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF4.1 GBneeds 4.9 GBRuns in the app today

    Scores

  • Qwen3.5-4B

    Qwen · packaged by unsloth

    A general assistant from Qwen's Qwen3.5 family with 4 billion parameters that reads images as well as text.

    Runs in the app today · Chat, and questions about images

    Parameters
    4B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 6.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.2 GBneeds 6.1 GBRuns in the app today

    Scores

    • MMLU-Pro79.1% (the original model's score)
    • GPQA Diamond76.2% (the original model's score)
    • IFEval89.8% (the original model's score)
    • LiveCodeBench v655.8% (the original model's score)
    • MMMU77.6% (the original model's score)
  • Gemma 4 E2B

    Google · packaged by ggml-org

    A small Gemma 4 model from Google, with 2.3 billion effective parameters.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    5.1B
    Context
    131,072 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 7.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF6.0 GBneeds 7.1 GBRuns in the app today

    Scores

  • Qwen3.5-9B

    Qwen · packaged by unsloth

    A general assistant from Qwen's Qwen3.5 family with 9 billion parameters that reads images as well as text.

    Runs in the app today · Chat, and questions about images

    Parameters
    9B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 7.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF6.6 GBneeds 7.8 GBRuns in the app today

    Scores

    • MMLU-Pro82.5% (the original model's score)
    • GPQA Diamond81.7% (the original model's score)
    • IFEval91.5% (the original model's score)
    • LiveCodeBench v665.6% (the original model's score)
    • MMMU78.4% (the original model's score)
  • NVIDIA Nemotron 3.5 Bolt E4B

    NVIDIA · packaged by ggml-org

    Nemotron 3.5 Bolt E4B is NVIDIA's further-trained version of Google's Gemma 4 E4B, with 8 billion parameters, 4 billion effective.

    Runs in the app today · Chat

    Parameters
    8B
    Context
    65,536 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 9.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF8.0 GBneeds 9.5 GBRuns in the app today

    Scores

  • Gemma 4 E4B

    Google · packaged by ggml-org

    A small Gemma 4 model from Google, with 4.5 billion effective parameters.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    8B
    Context
    131,072 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 10.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.0 GBneeds 10.7 GBRuns in the app today

    Scores

  • Granite-4.2-8B

    IBM

    An 8-billion-parameter chat model from IBM that reasons step by step before it answers, with modes for full, low-effort or no thinking.

    Runs in the app today · Chat

    Parameters
    8B
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 11.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.4 GBneeds 11.0 GBRuns in the app today

    Scores

  • Ministral 3 8B Instruct

    Mistral AI

    Ministral 3 8B Instruct is an 8-billion-parameter chat model from Mistral AI that also reads images.

    Runs in the app today · Chat, and questions about images

    Parameters
    8B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 11.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.9 GBneeds 11.7 GBRuns in the app today

    Scores

  • MiMo-V2.6-Distill-Qwen-9B

    Xiaomi · packaged by ggml-org

    MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter model from Xiaomi, made by fine-tuning Qwen3.5-9B on data from its larger MiMo models.

    Runs in the app today · Chat, and questions about images

    Parameters
    9B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 12.0 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF10.2 GBneeds 12.0 GBRuns in the app today

    Scores

General chat, medium

10 models, 10 in the app today

Assistants of 10 to 40 billion parameters. By the app's estimate, the smallest version of each needs 8.5 to 34.8 GB of GPU memory.

  • Gemma 4 12B

    Google

    A Gemma 4 model from Google with 12 billion parameters.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    11.9B
    Context
    262,144 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 8.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_0 GGUF7.2 GBneeds 8.5 GBRuns in the app today

    Scores

  • Ministral 3 14B Instruct

    Mistral AI

    Ministral 3 14B Instruct is a 14-billion-parameter chat model from Mistral AI that also reads images.

    Runs in the app today · Chat, and questions about images

    Parameters
    14B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 10.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF9.1 GBneeds 10.8 GBRuns in the app today

    Scores

  • gpt-oss-20b

    OpenAI · packaged by ggml-org

    gpt-oss-20b is the smaller of OpenAI's two open-weight reasoning models, a mixture-of-experts text model with 21 billion parameters, 3.6 billion active.

    Runs in the app today · Chat

    Parameters
    21B total, 3.6B active
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 14.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF12.1 GBneeds 14.3 GBRuns in the app today
  • Gemma 4 26B A4B

    Google

    A mixture-of-experts Gemma 4 model from Google with 25.2 billion parameters, 3.8 billion of them active for each token.

    Runs in the app today · Chat, and questions about images

    Parameters
    25.2B total, 3.8B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 18.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_0 GGUF15.6 GBneeds 18.4 GBRuns in the app today

    Scores

  • GLM-4.7-Flash

    Z.ai · packaged by ggml-org

    GLM-4.7-Flash is a mixture-of-experts text model from Z.ai with 30 billion parameters, 3 billion active per token.

    Runs in the app today · Chat

    Parameters
    30B total, 3B active
    Context
    202,752 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 21.5 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K GGUF18.2 GBneeds 21.5 GBRuns in the app today

    Scores

    • AIME 2591.6% (the original model's score)
    • GPQA75.2% (the original model's score)
    • LCB v664.0% (the original model's score)
    • HLE14.4% (the original model's score)
    • SWE-bench Verified59.2% (the original model's score)
  • Gemma 4 31B

    Google

    A dense Gemma 4 model from Google with 30.7 billion parameters.

    Runs in the app today · Chat, and questions about images

    Parameters
    30.7B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 22.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_0 GGUF18.9 GBneeds 22.2 GBRuns in the app today

    Scores

  • Qwen3.6-35B-A3B

    Qwen · packaged by ggml-org

    A mixture-of-experts model from Qwen's Qwen3.6 series with 35 billion parameters, 3 billion of them active for each token.

    Runs in the app today · Chat, and questions about images

    Parameters
    35B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 25.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF21.5 GBneeds 25.3 GBRuns in the app today

    Scores

  • Muse Glimmer 30B

    Meta · packaged by ggml-org

    Muse Glimmer is a 30-billion-parameter model from Meta that reads text and images, distilled from Muse Spark for agents.

    Runs in the app today · Chat, and questions about images

    Parameters
    29.6B
    Context
    131,072 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 27.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF23.1 GBneeds 27.2 GBRuns in the app today

    Scores

  • NVIDIA Nemotron 3.5 Lightning 30B-A3B

    NVIDIA · packaged by unsloth

    Nemotron 3.5 Lightning is a mixture-of-experts text model from NVIDIA with 30 billion parameters, 3 billion active, that mixes Mamba-2 and attention layers.

    Runs in the app today · Chat

    Parameters
    30B total, 3B active
    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 29.8 GB
    License
    OpenMDW License Agreement 1.1 (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_M GGUF25.3 GBneeds 29.8 GBRuns in the app today

    Scores

  • Qwen3.8-27B

    Qwen · packaged by ggml-org

    A dense model from Qwen's Qwen3.8 series with 27 billion parameters, for coding, research and multi-step agent tasks.

    Runs in the app today · Chat, and questions about images

    Parameters
    27B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 34.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF29.5 GBneeds 34.8 GBRuns in the app today

    Scores

    Measured by 6529 AI

    • 64 tokens a second · One H100, 80 GB · Nov 13, 2026It also described an image correctly.

General chat, large

5 models, 5 in the app today

Assistants of 40 to 150 billion parameters. By the app's estimate, the smallest version of each needs 74.6 to 132 GB of GPU memory.

  • gpt-oss-120b

    OpenAI · packaged by ggml-org

    gpt-oss-120b is the larger of OpenAI's two open-weight reasoning models, a mixture-of-experts text model with 117 billion parameters, 5.1 billion active.

    Runs in the app today · Chat

    Parameters
    117B total, 5.1B active
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 74.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF63.4 GBneeds 74.6 GBRuns in the app today

    Measured by 6529 AI

    • 120.6 tokens a second · 8 × A100, 40 GB each · Nov 14, 2026256 tokens at temperature 0 after a warm-up.
  • NVIDIA Nemotron 3.5 Super VL 120B-A12B

    NVIDIA · packaged by ggml-org

    Nemotron 3.5 Super VL is a mixture-of-experts model from NVIDIA with 120 billion parameters, 12 billion active, that reads text, images and video.

    Runs in the app today · Chat, and questions about images

    Parameters
    120B total, 12B active
    Context
    1,048,576 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 86.9 GB
    License
    OpenMDW License Agreement 1.1 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_0 GGUF73.8 GBneeds 86.9 GBRuns in the app today

    Scores

  • Mistral Small 4 119B

    Mistral AI · packaged by unsloth

    Mistral Small 4 is a mixture-of-experts model with 119 billion parameters, 6.5 billion active per token, that reads text and images.

    Runs in the app today · Chat, and questions about images

    Parameters
    119B total, 6.5B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 87.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_M GGUF74.6 GBneeds 87.8 GBRuns in the app today

    Scores

    • AA LCR72.0% (the original model's score)
  • Mistral Medium 3.5 128B

    Mistral AI · packaged by unsloth

    Mistral Medium 3.5 is a dense 128-billion-parameter model from Mistral AI that reads text and images and handles chat, reasoning and coding in one set of weights.

    Runs in the app today · Chat, and questions about images

    Parameters
    128B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 94.5 GB
    License
    Modified MIT License (conditional)
    License terms

    Every copy carries the license and its attribution notice. A provider whose company, or employer, had more than 20 million US dollars of revenue in the previous month has no rights under the license.

    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF80.3 GBneeds 94.5 GBRuns in the app today

    Scores

  • Qwen3.8-Flash-Next

    Qwen · packaged by unsloth

    An experimental preview from Qwen of the architecture planned for Qwen4: a mixture-of-experts model with 125 billion parameters, 6 billion of them active for each token.

    Runs in the app today · Chat, and questions about images

    Parameters
    125B total, 6B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 132.1 GB
    License
    Qwen Community License 1.0 (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service or AI work assistant business needs a separate license from Qwen before any commercial use. Internal use, which makes the model available to no third party, is exempt. A commercial product with more than 100 million monthly active users or 20 million US dollars of monthly revenue shows the model's name on its user interface.

    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF112.2 GBneeds 132.1 GBRuns in the app today

    Scores

General chat, very large

10 models, 10 in the app today

The largest open models, over 150 billion parameters. By the app's estimate, the smallest version of each needs 128.2 to 1,776 GB of GPU memory, so they need several data-center GPUs.

  • Step 3.7 Flash

    StepFun

    Step 3.7 Flash is a mixture-of-experts model from StepFun with 198 billion parameters, about 11 billion active, that reads text and images.

    Runs in the app today · Chat, and questions about images

    Parameters
    198B total, 11B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 128.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    IQ4_XS GGUF109.0 GBneeds 128.2 GBRuns in the app today

    Scores

  • MiMo-V2.6-Flash-RL

    Xiaomi · packaged by ggml-org

    MiMo-V2.6-Flash-RL is a mixture-of-experts model from Xiaomi with 309 billion parameters, 15 billion active, trained with reinforcement learning on coding, agent, visual and security tasks.

    Runs in the app today · Chat

    Parameters
    309B total, 15B active
    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 197.0 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF167.4 GBneeds 197.0 GBRuns in the app today

    Scores

  • GLM-5.3-Flash

    Z.ai · packaged by ggml-org

    GLM-5.3-Flash is the first GLM-5 model from Z.ai that reads images as well as text, with 320 billion parameters and 18 billion active per token.

    Runs in the app today · Chat, and questions about images

    Parameters
    320B total, 18B active
    Context
    1,048,576 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 214.2 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K GGUF182.0 GBneeds 214.2 GBRuns in the app today

    Measured by 6529 AI

    • 59.5 tokens a second · 8 × H100, 80 GB each · Nov 14, 2026256 tokens at temperature 0 after a warm-up. It also described a 512 by 512 image correctly.
  • MiniMax-M3

    MiniMax · packaged by unsloth

    MiniMax-M3 is a mixture-of-experts model with about 428 billion parameters, 23 billion active, trained on text, images and video together.

    Runs in the app today · Chat

    Parameters
    428B total, 23B active
    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 311.7 GB
    License
    MiniMax Community License (conditional)
    License terms

    The base grant is for non-commercial use. Every copy carries the license and its notice. Commercial use shows "Built with MiniMax M3" and sends MiniMax a one-time notice, at the address and with the subject line the license gives. Above 20 million US dollars of yearly revenue it needs MiniMax's written authorization first. The license lists prohibited uses, military use among them.

    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_M GGUF264.9 GBneeds 311.7 GBRuns in the app today
  • DeepSeek-V4.1-Flash-1017

    DeepSeek · packaged by unsloth

    DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, a model for reasoning, coding and agent tasks.

    Runs in the app today · Chat

    The app serves it on its chat route. 6529 replaced unsloth's eighth shard, whose engram table mainline llama.cpp reads wrong, with one made from unsloth's Q6_K copy of the same table. llama.cpp reads that 161 GB table from disk as needed, so in 6529's run the GPUs held 311 GB at a 32,768-token context, less than the app's estimate of 540.3 GB. These files have no image encoder, so the app gives it text only.

    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 540.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF459.2 GBneeds 540.3 GBRuns in the app today

    Scores

    Measured by 6529 AI

    • 64.5 to 64.8 tokens a second · 8 × H200, 141 GB each · Nov 16, 2026At temperature 0, on four replies of 103 to 486 tokens through the app's chat route, and every test answer was right. The GPUs held 296,272 MiB in all, and the 161 GB engram table stayed on disk.
  • GLM-5.3

    Z.ai · packaged by unsloth

    GLM-5.3 is a text model from Z.ai for coding, agent tasks that run for hours, and finding software vulnerabilities.

    Runs in the app today · Chat

    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 549.8 GB
    License
    GLM-5.3 License (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service business with more than 10 billion US dollars of revenue over any 12 months must pass Z.AI's security review before any commercial use.

    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF467.3 GBneeds 549.8 GBRuns in the app today

    Scores

  • MiMo-V2.6-Pro-RL

    Xiaomi · packaged by ggml-org

    MiMo-V2.6-Pro-RL is the larger MiMo-V2.6 model from Xiaomi, a mixture-of-experts model with 1.02 trillion parameters, 42 billion active.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    1.02T total, 42B active
    Context
    1,048,576 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 655.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF557.1 GBneeds 655.4 GBRuns in the app today

    Scores

  • Mistral Large 4

    Mistral AI · packaged by ggml-org

    Mistral Large 4 is a mixture-of-experts model with about 1 trillion parameters, 52 billion active, that reads text and images.

    Runs in the app today · Chat, and questions about images

    Parameters
    1T total, 52B active
    Context
    1,048,576 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 750.9 GB
    License
    Mistral Ecosystem Open-Weight License (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service business needs a commercial license from Mistral before any commercial use. Internal use, which makes the model available to no third party, is exempt.

    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF638.2 GBneeds 750.9 GBRuns in the app today

    Scores

  • DeepSeek-V4-Pro-0813

    DeepSeek · packaged by unsloth

    DeepSeek-V4-Pro-0813 is the full release of DeepSeek V4 Pro, a mixture-of-experts text model for reasoning, coding and agent tasks that run for many steps.

    Runs in the app today · Chat

    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 999.7 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF849.7 GBneeds 999.7 GBRuns in the app today

    Scores

  • Kimi K3

    Moonshot AI · packaged by unsloth

    Kimi K3 is a mixture-of-experts model from Moonshot AI with 2.8 trillion parameters, 104 billion active, that reads text, images and video.

    Runs in the app today · Chat, and questions about images

    Parameters
    2.8T total, 104B active
    Context
    1,048,576 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 1,776.0 GB
    License
    Kimi K3 License (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service business with more than 20 million US dollars of revenue over any 12 months needs an agreement with Moonshot AI before any commercial use. A commercial product with more than 100 million monthly active users or 20 million US dollars of monthly revenue shows "Kimi K3" on its user interface. Internal use is exempt from both.

    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF1,509.6 GBneeds 1,776.0 GBRuns in the app today

    Scores

Coding and math

5 models, 5 in the app today

Models trained to write and fix code, or to solve math problems.

  • Qwen2.5-Math-1.5B-Instruct

    Qwen · packaged by lmstudio-community

    A small math model from Qwen with 1.5 billion parameters that solves English and Chinese math problems step by step, or by writing code for a tool to run.

    Runs in the app today · Chat

    Parameters
    1.5B
    Context
    4,096 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.7 GBneeds 2.0 GBRuns in the app today

    Scores

    • MATH79.7% (the original model's score)
  • Qwen2.5-Coder-1.5B-Instruct

    Qwen

    A small coding assistant from Qwen's Qwen2.5-Coder family with 1.54 billion parameters.

    Runs in the app today · Chat

    Parameters
    1.5B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.9 GBneeds 2.3 GBRuns in the app today
  • Devstral Small 2 24B Instruct

    Mistral AI · packaged by unsloth

    Devstral Small 2 is a 24-billion-parameter coding model from Mistral AI that uses tools to explore codebases, edit several files and fix issues.

    Runs in the app today · Chat, and questions about images

    Parameters
    24B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 17.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF15.2 GBneeds 17.9 GBRuns in the app today

    Scores

  • Qwen3-Coder-30B-A3B-Instruct

    Qwen · packaged by ggml-org

    A coding model from Qwen's Qwen3-Coder family with 30.5 billion parameters, 3.3 billion of them active for each token.

    Runs in the app today · Chat

    Parameters
    30.5B total, 3.3B active
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 21.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF18.0 GBneeds 21.2 GBRuns in the app today
  • Qwen3-Coder-Next

    Qwen

    A coding model from Qwen made for coding agents and local development, with 80 billion parameters, 3 billion of them active for each token.

    Runs in the app today · Chat

    Parameters
    80B total, 3B active
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 57.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF48.4 GBneeds 57.0 GBRuns in the app today

    Measured by 6529 AI

    • 38 tokens a second · One H100, 80 GB, experts in system memory · Nov 13, 2026The config passed --n-cpu-moe 999.

Vision and multimodal

10 models, 10 in the app today

Models built to answer questions about images, video or audio as well as text.

  • SmolVLM2-500M-Video-Instruct

    Hugging Face · packaged by ggml-org

    A 500-million-parameter model from Hugging Face that answers questions about videos, images and text in English and reads text in images.

    Runs in the app today · Chat, and questions about images

    Parameters
    500M
    Context
    8,192 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today

    Scores

    • Video-MME42.2% (the original model's score)
    • MLVU47.3% (the original model's score)
    • MVBench39.7% (the original model's score)
  • MiniCPM-V 4.7 1B

    OpenBMB

    A 1-billion-parameter vision-language model from OpenBMB that answers questions about single images, several images and video.

    Runs in the app today · Chat, and questions about images

    Parameters
    1B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 2.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.2 GBneeds 2.6 GBRuns in the app today
  • LFM3-Omni-2B

    Liquid AI

    A 2.2-billion-parameter model from Liquid AI that takes text, images, video and 16 kHz audio and answers in text or 24 kHz speech.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    2.2B
    Context
    32,768 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 3.3 GB
    License
    LFM Open License v1.0 (conditional)
    License terms

    Every copy carries the license and its notices. Commercial use is licensed only to an entity with less than 10 million US dollars of annual revenue.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.8 GBneeds 3.3 GBRuns in the app today

    Scores

  • Qwen3-VL-4B-Instruct

    Qwen

    A vision-language model from Qwen with 4 billion parameters that answers questions about images and video, reads text in 32 languages, and can operate computer and phone screens as an agent.

    Runs in the app today · Chat, and questions about images

    Parameters
    4B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 6.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.1 GBneeds 6.1 GBRuns in the app today
  • Qwen3-VL-8B-Instruct

    Qwen

    A vision-language model from Qwen with 8 billion parameters that answers questions about images and video, reads text in 32 languages, and can operate computer and phone screens as an agent.

    Runs in the app today · Chat, and questions about images

    Parameters
    8B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 11.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.9 GBneeds 11.7 GBRuns in the app today

    Measured by 6529 AI

    • 78 tokens a second · One L40S, 48 GB · Nov 14, 2026
  • Qwen3-Omni-30B-A3B-Instruct

    Qwen · packaged by ggml-org

    A mixture-of-experts model from Qwen that understands text, images, audio and video, with 30 billion parameters and 3 billion active for each token.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    30B total, 3B active
    Context
    65,536 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 24.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF20.8 GBneeds 24.5 GBRuns in the app today

    Scores

    • GPQA69.6% (the original model's score)
    • AIME2565.0% (the original model's score)
    • MMMU_val69.1% (the original model's score)
    • Video-MME70.5% (the original model's score)
    • LibriSpeech clean1.2% (the original model's score)lower is better
  • Qwen3-Omni-30B-A3B-Thinking

    Qwen · packaged by ggml-org

    The reasoning version of Qwen3-Omni: a mixture-of-experts model from Qwen that reads text, images, audio and video and thinks step by step before answering in text.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    30B total, 3B active
    Context
    65,536 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 24.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF20.8 GBneeds 24.5 GBRuns in the app today

    Scores

    • GPQA73.1% (the original model's score)
    • AIME2573.7% (the original model's score)
    • Arena-Hard v255.1% (the original model's score)
    • MMMU_val75.6% (the original model's score)
    • Video-MME69.7% (the original model's score)
  • Qwen3-VL-32B-Instruct

    Qwen

    A vision-language model from Qwen with 32 billion parameters that answers questions about images and video, reads text in 32 languages, and can operate computer and phone screens as an agent.

    Runs in the app today · Chat, and questions about images

    Parameters
    32B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 24.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF21.0 GBneeds 24.7 GBRuns in the app today
  • MiniCPM-V 4.7 35B-A3B

    OpenBMB

    A mixture-of-experts vision-language model from OpenBMB with 35 billion parameters, 3 billion of them active per token.

    Runs in the app today · Chat, and questions about images

    Parameters
    35B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 26.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF22.3 GBneeds 26.3 GBRuns in the app today
  • NVIDIA Nemotron 3 Nano Omni 30B-A3B Reasoning

    NVIDIA · packaged by ggml-org

    Nemotron 3 Nano Omni is a mixture-of-experts model from NVIDIA with 31 billion parameters, about 3 billion active, that reads video, audio, images and text.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    31B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 29.9 GB
    License
    NVIDIA Open Model Agreement (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF25.4 GBneeds 29.9 GBRuns in the app today

    Scores

Documents and OCR

24 models, 15 in the app today

Models that read scanned pages, PDFs, tables and charts into text, Markdown or structured data.

  • Table Transformer (table detection)

    Microsoft

    Table Transformer finds the tables in an image of a document page and draws a box around each one.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.1 GBneeds 0.2 GBNot yet
  • Table Transformer (table structure recognition)

    Microsoft

    This Table Transformer finds the rows, columns and other parts of a table in an image of that table.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.1 GBneeds 0.2 GBNot yet
  • Granite Docling 258M

    IBM · packaged by ggml-org

    Granite Docling turns a document page into structured text for the Docling library, which exports it as Markdown or HTML with tables, code and equations.

    Runs in the app today · Chat, and questions about images

    Parameters
    258M
    Context
    8,192 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 0.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.4 GBneeds 0.5 GBRuns in the app today

    Scores

  • OvisOCR2

    ATH-MaaS

    OvisOCR2 turns a document page image into Markdown in reading order, with formulas as LaTeX, tables as HTML and boxes around pictures and charts.

    Runs in the app today · Chat, and questions about images

    Parameters
    800M
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.0 GBneeds 1.2 GBRuns in the app today
    BF16 GGUF1.7 GBneeds 2.1 GBRuns in the app today
    safetensors1.7 GBneeds 2.1 GBNot yet
    Its engine, PyTorch, is not in the app.
  • DePlot

    Google

    DePlot reads an image of a chart or plot and writes out the data table behind it.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.1 GBneeds 1.4 GBNot yet
  • GOT-OCR 2.0

    StepFun

    GOT-OCR 2.0 reads text from documents, scene photos, tables, charts, math and molecular formulas, geometric shapes and sheet music.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.1 GBneeds 1.4 GBNot yet
  • GLM-OCR

    Z.ai · packaged by ggml-org

    GLM-OCR reads complex documents, including tables, formulas, code and seals, and fills a JSON template with fields it finds, such as the details on an ID card.

    Runs in the app today · Chat, and questions about images

    Parameters
    900M
    Context
    131,072 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.7 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.4 GBneeds 1.7 GBRuns in the app today

    Scores

  • LightOnOCR-2-1B

    LightOn · packaged by ggml-org

    LightOnOCR-2 converts PDFs, scans and document images into clean text in reading order, including tables, receipts, forms, multi-column pages and math.

    Runs in the app today · Chat, and questions about images

    Parameters
    1B
    Context
    16,384 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.5 GBneeds 1.8 GBRuns in the app today
  • HunyuanOCR

    Tencent · packaged by ggml-org

    HunyuanOCR is a small vision-language model for reading text in images.

    Runs in the app today · Chat, and questions about images

    Reads
    Text, images
    Makes
    Text
    Memory
    from 1.9 GB
    License
    Tencent Hunyuan Community License Agreement (conditional)
    License terms

    The license does not apply in the European Union, the United Kingdom or South Korea, for serving or for pinning. Every copy carries the license, and a distribution carries a Notice file with the notice the license gives. Every agreement with a user must include the use restrictions and the acceptable use policy. A licensee whose products had more than 100 million monthly active users on the release date needs a license from Tencent. The license doesn't apply in EU, GB, KR.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.6 GBneeds 1.9 GBRuns in the app today
  • MiniCPM-OCR 1.0

    OpenBMB

    MiniCPM-OCR 1.0 parses Chinese and English documents and PDFs into text, tables, formulas, charts and seals, and extracts key fields.

    Runs in the app today · Chat, and questions about images

    Parameters
    1.3B
    Reads
    Text, images
    Makes
    Text
    Memory
    from 2.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.7 GBneeds 2.1 GBRuns in the app today
  • MinerU2.5-Pro-2605-1.2B

    OpenDataLab

    MinerU2.5-Pro converts Chinese and English PDF pages to Markdown, with text, formulas, tables, images and charts in reading order.

    Runs in the app today · Chat, and questions about images

    Parameters
    1.2B
    Context
    32,768 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 2.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.9 GBneeds 2.2 GBRuns in the app today
    BF16 GGUF2.3 GBneeds 2.8 GBRuns in the app today
    safetensors2.3 GBneeds 2.8 GBNot yet
    Its engine, PyTorch, is not in the app.
  • PaddleOCR-VL-1.6

    PaddlePaddle

    PaddleOCR-VL-1.6 parses document pages into text, tables, formulas, charts and seals, and finds text with its position on the page.

    Runs in the app today · Chat, and questions about images

    Parameters
    900M
    Context
    131,072 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 2.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    GGUF1.8 GBneeds 2.2 GBRuns in the app today
  • TeleOCR

    XingChen-AGI

    TeleOCR parses digital documents and photos of documents, including curved or warped pages, into text, formulas and tables.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    1.2B
    Context
    128,000 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 3.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors2.9 GBneeds 3.4 GBNot yet
  • DeepSeek-OCR

    DeepSeek · packaged by ggml-org

    DeepSeek-OCR converts a document image to Markdown or plain text, with layout boxes on request.

    Runs in the app today · Chat, and questions about images

    Context
    8,192 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 4.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF3.6 GBneeds 4.3 GBRuns in the app today
  • NVIDIA Nemotron Parse 2.0

    NVIDIA

    Nemotron Parse 2.0 turns document images into structured text with layout classes, bounding boxes and reading order, including tables and charts.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    9,000 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 4.3 GB
    License
    OpenMDW License Agreement 1.1 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors3.6 GBneeds 4.3 GBNot yet
  • Unlimited-OCR

    Baidu

    Unlimited-OCR parses a document image, or a whole multi-page PDF in one pass, into structured text.

    Runs in the app today · Chat, and questions about images

    Context
    32,768 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 4.7 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF4.0 GBneeds 4.7 GBRuns in the app today
    BF16 GGUF6.7 GBneeds 7.9 GBRuns in the app today
    safetensors6.7 GBneeds 7.9 GBNot yet
    Its engine, PyTorch, is not in the app.
  • dots.mocr

    rednote hilab

    dots.mocr parses document pages in many languages into text with layout, reading order and tables.

    Runs in the app today · Chat, and questions about images

    Parameters
    3B
    Context
    131,072 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 5.2 GB
    License
    dots.ocr License Agreement (conditional)
    License terms

    Every copy carries the MIT license and the dots license agreement with their notices. The agreement lists prohibited uses, and a breach of them ends the license.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF4.4 GBneeds 5.2 GBRuns in the app today
    BF16 GGUF6.1 GBneeds 7.2 GBRuns in the app today
    safetensors6.1 GBneeds 7.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • dots.ocr

    rednote hilab · packaged by ggml-org

    dots.ocr parses document pages into text, tables and formulas with their layout and reading order, all in one vision-language model.

    Runs in the app today · Chat, and questions about images

    Parameters
    3B
    Context
    131,072 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 5.3 GB
    License
    dots.ocr License Agreement (conditional)
    License terms

    Every copy carries the MIT license and the dots license agreement with their notices. The agreement lists prohibited uses, and a breach of them ends the license.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF4.4 GBneeds 5.3 GBRuns in the app today

    Scores

  • NuExtract3

    NuMind

    A 4-billion-parameter vision-language model from NuMind that pulls fields from documents into JSON following a template, and converts page images to Markdown.

    Runs in the app today · Chat, and questions about images

    Parameters
    4B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 6.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.3 GBneeds 6.3 GBRuns in the app today

    Scores

  • Qianfan-OCR

    Baidu · packaged by ggml-org

    Qianfan-OCR converts document images straight to Markdown, answers questions about documents and charts, and pulls out key fields.

    Runs in the app today · Chat, and questions about images

    Parameters
    4B
    Context
    32,768 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 6.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.4 GBneeds 6.3 GBRuns in the app today

    Scores

  • DeepSeek-OCR 2

    DeepSeek

    DeepSeek-OCR 2 converts document images to Markdown or plain text, with layout boxes on request.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    8,192 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 8.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors6.8 GBneeds 8.0 GBNot yet
  • LightOnOCR-3-4B

    LightOn

    LightOnOCR-3-4B transcribes document pages to Markdown and, on request, labels each block with its type and a bounding box.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    4B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 10.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors9.1 GBneeds 10.8 GBNot yet
  • olmOCR-2-7B-1025

    Ai2 · packaged by lmstudio-community

    olmOCR 2 converts PDF pages and document images into text in reading order.

    Runs in the app today · Chat, and questions about images

    Parameters
    7B
    Context
    128,000 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 11.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.5 GBneeds 11.2 GBRuns in the app today

    Scores

  • Chandra OCR 2

    Datalab

    Chandra OCR 2 turns images and PDFs into Markdown, HTML or JSON and keeps the page layout.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 12.5 GB
    License
    AI Pubs Open RAIL-M License (Modified) (conditional)
    License terms

    Every copy carries the license, and its use restrictions bind every user. Except for personal use or research, the model may not be used by an entity with more than 2 million US dollars of gross revenue in the prior year, or more than 2 million US dollars of equity or debt funding raised in total, or by one that offers a product competing with the licensor's.

    Versions
    VersionFilesMemoryIn the app
    safetensors10.6 GBneeds 12.5 GBNot yet

Security research

9 models, 5 in the app today

Models for offensive and defensive security work. Those marked as refusal-removed answer requests their originals decline.

  • deberta-v3-base-prompt-injection-v2

    Protect AI

    A small classifier from Protect AI, fine-tuned from Microsoft's DeBERTa-v3-base, that labels English text as benign or as a prompt injection, an attempt to override a language model's instructions.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    512 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 0.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.8 GBneeds 0.9 GBNot yet

    Scores

  • Foundation-Sec-8B-Reasoning

    Foundation AI at Cisco

    An 8-billion-parameter security assistant from Foundation AI at Cisco, built on Llama 3.1 8B, that reasons step by step before it answers.

    Runs in the app today · Chat

    Parameters
    8B
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 10.1 GB
    License
    Apache License 2.0 and Llama 3.1 Community License (conditional)
    License terms

    Every copy carries both licenses and the Llama 3.1 notice. A product that contains the model shows "Built with Llama". A licensee whose products had more than 700 million monthly active users in the month before the Llama 3.1 release needs a license from Meta. Llama's acceptable use policy lists prohibited uses.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF8.5 GBneeds 10.1 GBRuns in the app today
    safetensors16.1 GBneeds 19.0 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • DeepHat-V1-7B

    DeepHat

    A 7.61-billion-parameter model fine-tuned from Qwen2.5-Coder-7B for offensive and defensive cybersecurity work and DevOps questions.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    7.6B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 18.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors15.3 GBneeds 18.0 GBNot yet
  • Qwen3.8-27B-Uncensored-Cyber

    philbert440

    Made from Qwen3.8-27B, Qwen's 27-billion-parameter assistant that reads text and images.

    Its refusals were removed and a cybersecurity refusal direction was removed on top, so it answers offensive-security requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    27B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 23.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF19.9 GBneeds 23.5 GBRuns in the app today
    Q8_0 GGUF29.5 GBneeds 34.8 GBRuns in the app today
    safetensors55.6 GBneeds 65.5 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • Huihui-CyberStrike-OffSec-35B-abliterated

    huihui-ai

    Made from CyberStrike-OffSec-35B, a Qwen3.6-35B-A3B fine-tune that emits tool calls for a penetration-testing agent, with 35 billion parameters and 3 billion active per token.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    35B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 26.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K GGUF22.6 GBneeds 26.7 GBRuns in the app today
    Q8_0 GGUF38.7 GBneeds 45.6 GBRuns in the app today
    safetensors71.9 GBneeds 84.7 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Cyber-Tiel-Coder-35B-A3B

    peculiar-ragdoll

    Made from Ornith-1.5-35B-A3B, a mixture-of-experts coding and agent model with 35 billion parameters, 3 billion active per token, that also reads images.

    Its refusals were removed, so it answers requests the original declines, offensive-security work among them.

    Runs in the app today · Chat, and questions about images

    Parameters
    35B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 27.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF23.3 GBneeds 27.4 GBRuns in the app today
    UD-Q8_K_XL GGUF39.4 GBneeds 46.3 GBRuns in the app today

    Scores

  • Qwen3.8-Flash-Next-CYBERSECURITY

    dealignai

    Made from Qwen3.8-Flash-Next, Qwen's mixture-of-experts assistant with 125 billion parameters, 6 billion active per token, that reads text, images and video.

    Its refusals were removed by a direct weight edit aimed at offensive-security research, so it answers requests the original declines.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    125B total, 6B active
    Context
    262,144 tokens
    Reads
    Text, images, video
    Makes
    Text
    Memory
    from 159.2 GB
    License
    Qwen Community License 1.0 (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service or AI work assistant business needs a separate license from Qwen before any commercial use. Internal use, which makes the model available to no third party, is exempt. A commercial product with more than 100 million monthly active users or 20 million US dollars of monthly revenue shows the model's name on its user interface.

    Versions
    VersionFilesMemoryIn the app
    NVFP4 safetensors135.3 GBneeds 159.2 GBNot yet

    Scores

  • GLM-5.3-Flash-CYBERSECURITY

    dealignai · packaged by SyndicateLabs

    GLM-5.3-Flash-CYBERSECURITY is made from Z.ai's GLM-5.3-Flash by dealignai.

    Its refusals were removed, so it answers offensive-security and other harmful requests the original declines.

    Runs in the app today · Chat

    Parameters
    320B total, 18B active
    Context
    1,048,576 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 229.1 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF195.5 GBneeds 230.1 GBRuns in the app today
    W4A16 safetensors194.7 GBneeds 229.1 GBNot yet
    Its engine, PyTorch, is not in the app.
  • GLM-5.3-CYBERSECURITY

    dealignai

    GLM-5.3-CYBERSECURITY is made from Z.ai's GLM-5.3 by dealignai.

    Its refusals were removed for offensive-security work, so it answers attack, exploit and malware requests the original declines.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    753B
    Context
    1,048,576 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 889.1 GB
    License
    GLM-5.3 License (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service business with more than 10 billion US dollars of revenue over any 12 months must pass Z.AI's security review before any commercial use.

    Versions
    VersionFilesMemoryIn the app
    FP8 safetensors755.7 GBneeds 889.1 GBNot yet

Refusals removed

12 models, 11 in the app today

General models whose publishers removed the refusals, so they answer requests the originals decline.

  • Gemma-4-E4B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    Made from Google's Gemma 4 E4B, a small assistant with 4.5 billion effective parameters that reads text, images and audio.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    The chat page has no audio input, so the page can't use this model's audio input today.

    Parameters
    8B
    Context
    131,072 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 7.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF6.3 GBneeds 7.5 GBRuns in the app today
    Q8_K_P GGUF9.1 GBneeds 10.8 GBRuns in the app today

    Scores

  • Qwen3.5-9B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    Made from Qwen3.5-9B, Qwen's 9-billion-parameter assistant that reads text, images and video in 201 languages and dialects.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    9B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 7.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF6.6 GBneeds 7.8 GBRuns in the app today
    Q8_0 GGUF10.5 GBneeds 12.3 GBRuns in the app today

    Scores

  • Qwen-Image-2.1-Turbo-Uncensored

    AtomicChat

    Made from Qwen3-VL-8B-Instruct, the 8-billion-parameter text encoder of the Qwen-Image 2.1 image generator, which also runs as a chat model that reads images.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    The app serves it on its chat route, as a model that reads images. It doesn't pair it with an image generator.

    Parameters
    8B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 11.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.9 GBneeds 11.7 GBRuns in the app today
    BF16 GGUF17.6 GBneeds 20.7 GBRuns in the app today
  • Huihui-gpt-oss-20b-abliterated-v2

    huihui-ai

    Huihui-gpt-oss-20b-abliterated-v2 is made from OpenAI's gpt-oss-20b by huihui-ai.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat

    Parameters
    21B total, 3.6B active
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 16.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF13.8 GBneeds 16.3 GBRuns in the app today
    MXFP4 safetensors13.8 GBneeds 16.3 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Huihui-Qwen3.8-27B-abliterated

    huihui-ai

    Made from Qwen3.8-27B, Qwen's 27-billion-parameter assistant that reads text, images and video and works through coding and agent tasks.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    27B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 21.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF18.3 GBneeds 21.6 GBRuns in the app today
    UD-Q8_K_XL GGUF32.4 GBneeds 38.2 GBRuns in the app today
    safetensors55.6 GBneeds 65.4 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • Qwen3.8-27B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    Made from Qwen3.8-27B, Qwen's 27-billion-parameter assistant that reads text, images and video and works through coding and agent tasks.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    27B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 23.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_P GGUF19.8 GBneeds 23.3 GBRuns in the app today
    Q8_K_P GGUF33.3 GBneeds 39.2 GBRuns in the app today

    Scores

  • gemma-4-31B-it-uncensored-heretic

    llmfan46

    Made from Google's Gemma 4 31B, a 30.7-billion-parameter assistant that reads text and images in over 140 languages.

    Its refusals were removed, so it answers most requests the original declines. llmfan46 counts 10 refusals in 100 test prompts, against 99 for the original.

    Runs in the app today · Chat, and questions about images

    Parameters
    30.7B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 23.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF19.9 GBneeds 23.4 GBRuns in the app today
    Q8_0 GGUF33.8 GBneeds 39.9 GBRuns in the app today

    Scores

  • Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    Made from Qwen3.6-35B-A3B, Qwen's mixture-of-experts assistant with 35 billion parameters, 3 billion active per token, that reads text, images and video.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    35B total, 3B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 26.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF22.1 GBneeds 26.0 GBRuns in the app today
    Q8_K_P GGUF44.5 GBneeds 52.4 GBRuns in the app today

    Scores

  • GPTOSS-120B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    GPTOSS-120B-Uncensored-HauhauCS-Aggressive is made from OpenAI's gpt-oss-120b.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat

    Parameters
    117B total, 5.1B active
    Context
    131,072 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 77.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    MXFP4 GGUF65.4 GBneeds 77.0 GBRuns in the app today
  • Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive

    HauhauCS

    Made from Qwen3.5-122B-A10B, Qwen's mixture-of-experts assistant with 122 billion parameters, 10 billion active per token, that reads text, images and video.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    122B total, 10B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 88.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF75.1 GBneeds 88.4 GBRuns in the app today
    Q8_K_P GGUF145.4 GBneeds 171.1 GBRuns in the app today

    Scores

  • Huihui-Qwen3.8-Flash-Next-abliterated-v2

    huihui-ai

    Made from Qwen3.8-Flash-Next, Qwen's mixture-of-experts assistant with 125 billion parameters, 6 billion active per token, that reads text and images.

    Its refusals were removed, so it answers requests the original declines.

    Runs in the app today · Chat, and questions about images

    Parameters
    125B total, 6B active
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 132.1 GB
    License
    Qwen Community License 1.0 (conditional)
    License terms

    Every copy carries the license and its notice. A provider that runs a Model-as-a-Service or AI work assistant business needs a separate license from Qwen before any commercial use. Internal use, which makes the model available to no third party, is exempt. A commercial product with more than 100 million monthly active users or 20 million US dollars of monthly revenue shows the model's name on its user interface.

    Versions
    VersionFilesMemoryIn the app
    UD-Q4_K_XL GGUF112.2 GBneeds 132.1 GBRuns in the app today
    Q8_0 GGUF189.1 GBneeds 222.6 GBRuns in the app today

    Scores

  • DeepSeek-V4.1-Flash-1017-UNCENSORED

    dealignai

    DeepSeek-V4.1-Flash-1017-UNCENSORED is made from DeepSeek-V4.1-Flash-1017 by dealignai.

    Its refusals were removed, so it answers requests the original declines.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    1,048,576 tokens
    Reads
    Text, images, video
    Makes
    Text
    Memory
    from 600.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    FP8 safetensors510.3 GBneeds 600.4 GBNot yet

Speech to text

14 models, 9 in the app today

Models that transcribe speech, some with timestamps, speaker labels or translation.

  • Whisper tiny

    OpenAI

    The 39-million-parameter size of OpenAI's Whisper, for speech-to-text and translation into English.

    Runs in the app today · Speech to text

    Parameters
    39M
    Reads
    Audio
    Makes
    Text
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    F16 GGML78 MBneeds 0.1 GBRuns in the app today
    safetensors0.2 GBneeds 0.2 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • Whisper base

    OpenAI · packaged by ggerganov

    OpenAI's Whisper base turns speech into text and translates speech in other languages into English.

    Runs in the app today · Speech to text

    Parameters
    74M
    Reads
    Audio
    Makes
    Text
    Memory
    from 0.2 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    GGML0.2 GBneeds 0.2 GBRuns in the app today

    Scores

  • Wav2Vec2-Base-960h

    Meta

    Meta's Wav2Vec2 base model, fine-tuned on 960 hours of LibriSpeech to transcribe English speech.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Audio
    Makes
    Text
    Memory
    from 0.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.4 GBneeds 0.5 GBNot yet

    Scores

  • Whisper large-v3 turbo

    OpenAI · packaged by ggerganov

    A faster version of OpenAI's Whisper large-v3 for speech-to-text and translation into English.

    Runs in the app today · Speech to text

    Parameters
    809M
    Reads
    Audio
    Makes
    Text
    Memory
    from 0.7 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q5_0 GGML0.6 GBneeds 0.7 GBRuns in the app today

    Measured by 6529 AI

    • 11 seconds of speech transcribed in 0.6 seconds · One H100, 80 GB · Nov 13, 2026
    • Ran a test of speech to text. · One L40S, 48 GB · Nov 14, 2026
  • Parakeet TDT 0.6B v3

    NVIDIA · packaged by ggml-org

    NVIDIA's 600-million-parameter speech recognizer for 25 European languages.

    Runs in the app today · Speech to text

    Parameters
    600M
    Reads
    Audio
    Makes
    Text
    Memory
    from 0.8 GB
    License
    CC-BY-4.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGML0.7 GBneeds 0.8 GBRuns in the app today

    Scores

  • Qwen3-ASR-0.6B

    Qwen · packaged by ggml-org

    Qwen's 0.6B speech recognizer for 30 languages and 22 Chinese dialects, including singing and songs with background music.

    Runs in the app today · Chat

    The app serves it on its chat route. The chat page has no audio input, and the Speech screen lists whisper.cpp models, so the page can't transcribe with it today.

    Parameters
    600M
    Context
    65,536 tokens
    Reads
    Text, audio
    Makes
    Text
    Memory
    from 1.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.2 GBneeds 1.4 GBRuns in the app today

    Scores

  • MOSS-Transcribe-Diarize 0.9B

    OpenMOSS

    OpenMOSS's 0.9B model that turns long recordings with several speakers into one transcript with timestamps and speaker labels.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    900M
    Reads
    Audio
    Makes
    Text
    Memory
    from 2.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.8 GBneeds 2.2 GBNot yet

    Scores

  • Nemotron 3.5 ASR Streaming 0.6B

    NVIDIA

    NVIDIA's 600-million-parameter streaming speech recognizer for 40 language-locales, with punctuation and capitals.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    600M
    Reads
    Audio
    Makes
    Text
    Memory
    from 3.1 GB
    License
    OpenMDW License Agreement 1.1 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors2.6 GBneeds 3.1 GBNot yet
  • Qwen3-ASR-1.7B

    Qwen · packaged by ggml-org

    Qwen's 1.7B speech recognizer for 30 languages and 22 Chinese dialects, including singing and songs with background music.

    Runs in the app today · Chat

    The app serves it on its chat route. The chat page has no audio input, and the Speech screen lists whisper.cpp models, so the page can't transcribe with it today.

    Parameters
    1.7B
    Context
    65,536 tokens
    Reads
    Text, audio
    Makes
    Text
    Memory
    from 3.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.8 GBneeds 3.4 GBRuns in the app today

    Scores

  • Granite-Speech-4.1-2B-Plus

    IBM

    IBM's 2B speech recognizer for English, French, German, Spanish and Portuguese that can also label speakers and give word-level timestamps.

    Runs in the app today · Chat

    The app serves it on its chat route. The chat page has no audio input, and the Speech screen lists whisper.cpp models, so the page can't transcribe with it today.

    Parameters
    2B
    Context
    4,096 tokens
    Reads
    Text, audio
    Makes
    Text
    Memory
    from 3.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.9 GBneeds 3.5 GBRuns in the app today

    Scores

  • Whisper large-v3

    OpenAI

    OpenAI's Whisper large-v3 turns speech into text in many languages and translates it into English.

    Runs in the app today · Speech to text

    Parameters
    1.6B
    Reads
    Audio
    Makes
    Text
    Memory
    from 3.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    F16 GGML3.1 GBneeds 3.7 GBRuns in the app today
    safetensors3.1 GBneeds 3.7 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Voxtral Mini 4B Realtime 2602

    Mistral AI · packaged by ggml-org

    Mistral AI's 4B speech-to-text model that transcribes live audio in 13 languages as it streams.

    Runs in the app today · Chat

    The app serves it on its chat route. The chat page has no audio input, and the Speech screen lists whisper.cpp models, so the page can't transcribe with it today.

    Parameters
    4B
    Context
    131,072 tokens
    Reads
    Text, audio
    Makes
    Text
    Memory
    from 6.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.7 GBneeds 6.7 GBRuns in the app today

    Scores

  • ARK-ASR-3B

    AutoArk

    A 3B speech recognizer for 19 languages, among them Chinese, English, Japanese, Korean and most European Union languages.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    3B
    Reads
    Audio
    Makes
    Text
    Memory
    from 9.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors8.1 GBneeds 9.6 GBNot yet

    Scores

  • VibeVoice-ASR

    Microsoft

    Microsoft's speech recognizer that transcribes up to 60 minutes of audio in one pass and says who spoke when.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Audio
    Makes
    Text
    Memory
    from 20.5 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors17.4 GBneeds 20.5 GBNot yet

Text to speech

10 models, none in the app yet

Models that read text aloud, some in a voice cloned from a short sample.

  • Chatterbox-Turbo

    Resemble AI · packaged by ggml-org

    Resemble AI's 350M English text-to-speech model for voice agents that need a short delay, with voice cloning from a reference clip.

    Not in the app yet · The app has no route for speech.

    Parameters
    350M
    Reads
    Text, audio
    Makes
    Speech
    Memory
    from 1.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.0 GBneeds 1.3 GBNot yet
  • Chatterbox Multilingual V3

    Resemble AI · packaged by ggml-org

    Resemble AI's 0.5B text-to-speech model for 23 languages that can clone a voice from a reference clip.

    Not in the app yet · The app has no route for speech.

    Parameters
    500M
    Reads
    Text, audio
    Makes
    Speech
    Memory
    from 1.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.2 GBneeds 1.4 GBNot yet
  • VibeVoice-Realtime-0.5B

    Microsoft

    Microsoft's 0.5B research text-to-speech model that starts speaking about 300 ms after text begins to arrive.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    500M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 2.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors2.0 GBneeds 2.4 GBNot yet
  • Qwen3-TTS-12Hz-1.7B-Base

    Qwen · packaged by ggml-org

    Qwen's 1.7B text-to-speech model for 10 languages that clones a voice from 3 seconds of reference audio.

    Not in the app yet · The app has no route for speech.

    Parameters
    1.7B
    Context
    32,768 tokens
    Reads
    Text, audio
    Makes
    Speech
    Memory
    from 3.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.5 GBneeds 3.0 GBNot yet

    Scores

  • Ming-3.0-tts-0.5B

    inclusionAI

    inclusionAI's 0.5B text-to-speech model for 7 languages plus Sichuanese and Cantonese, with voice cloning and voices designed from a written description.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    500M
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 3.6 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors3.0 GBneeds 3.6 GBNot yet
  • Qwen3-TTS-12Hz-1.7B-CustomVoice

    Qwen

    Qwen's 1.7B text-to-speech model with 9 built-in voices that speak 10 languages.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    1.7B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 5.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors4.5 GBneeds 5.4 GBNot yet
  • Qwen3-TTS-12Hz-1.7B-VoiceDesign

    Qwen

    Qwen's 1.7B text-to-speech model that creates a voice from a written description and speaks with it in 10 languages.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    1.7B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 5.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors4.5 GBneeds 5.4 GBNot yet
  • VibeVoice-1.5B

    Microsoft

    Microsoft's research model that turns a script into up to 90 minutes of conversation, such as a podcast, with up to 4 speakers.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    1.5B
    Context
    65,536 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 6.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors5.4 GBneeds 6.4 GBNot yet
  • MOSS-TTS-Local-Transformer-v1.5

    OpenMOSS

    OpenMOSS's text-to-speech model for 31 languages with voice cloning, long-form speech, duration control and inline pause markers.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    32,768 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 10.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors9.1 GBneeds 10.8 GBNot yet
  • MOSS-TTS-v1.5

    OpenMOSS

    OpenMOSS's text-to-speech model for 31 languages with voice cloning, long-form speech, duration control and Pinyin or IPA pronunciation control.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    40,960 tokens
    Reads
    Text
    Makes
    Speech
    Memory
    from 20.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors17.0 GBneeds 20.0 GBNot yet

Image generation

13 models, 13 in the app today

Models that make and edit images from a text prompt.

  • Illustrious XL v2.0

    Onoma AI

    Illustrious XL v2.0 is a text-to-image model from Onoma AI.

    Runs in the app today · Images from a prompt

    Reads
    Text
    Makes
    Images
    Memory
    from 8.2 GB
    License
    CreativeML-OpenRAIL-M (conditional)
    License terms

    Every copy carries the license. Its use restrictions bind every user, and an agreement with a user must include them.

    Versions
    VersionFilesMemoryIn the app
    safetensors6.9 GBneeds 8.2 GBRuns in the app today
  • SDXL 1.0 base

    Stability AI

    SDXL 1.0 base is Stability AI's latent diffusion model that makes and modifies images from text prompts.

    Runs in the app today · Images from a prompt

    Reads
    Text
    Makes
    Images
    Memory
    from 8.2 GB
    License
    CreativeML Open RAIL++-M License (conditional)
    License terms

    Every copy carries the license. Its use restrictions bind every user, and an agreement with a user must include them.

    Versions
    VersionFilesMemoryIn the app
    safetensors6.9 GBneeds 8.2 GBRuns in the app today
  • FLUX.2 [klein] 4B

    Black Forest Labs · packaged by Comfy-Org

    FLUX.2 [klein] 4B is a 4 billion parameter model from Black Forest Labs that makes images from text and edits images using one or more reference pictures.

    Runs in the app today · Images from a prompt

    Parameters
    4B
    Reads
    Text, images
    Makes
    Images
    Memory
    from 19.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors16.1 GBneeds 19.0 GBRuns in the app today
  • MOSS-Image-SFT

    OpenMOSS · packaged by Comfy-Org

    MOSS-Image-SFT is a text-to-image model from OpenMOSS, fine-tuned from MOSS-Image-Base, for English and Chinese prompts.

    Runs in the app today · Images from a prompt

    Reads
    Text
    Makes
    Images
    Memory
    from 19.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors16.2 GBneeds 19.1 GBRuns in the app today
  • HiDream-O1-Image

    HiDream-ai · packaged by Comfy-Org

    HiDream-O1-Image is an 8B model that makes images from text at up to 2,048 by 2,048 pixels, renders long text, edits pictures and places a subject from reference photos into new scenes.

    Runs in the app today · Images from a prompt

    Parameters
    8B
    Reads
    Text
    Makes
    Images
    Memory
    from 19.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors16.4 GBneeds 19.3 GBRuns in the app today

    Scores

    • GenEval90.0% (the original model's score)
    • DPG-Bench89.8 (the original model's score)
    • HPSv310.4 (the original model's score)
    • CVTG-2K0.913 (the original model's score)
    • LongText-Bench-EN0.979 (the original model's score)
  • HiDream-O1-Image-Dev

    HiDream-ai · packaged by Comfy-Org

    HiDream-O1-Image-Dev is the distilled version of HiDream-O1-Image, an 8B model that makes and edits images at up to 2,048 by 2,048 pixels.

    Runs in the app today · Images from a prompt

    Parameters
    8B
    Reads
    Text
    Makes
    Images
    Memory
    from 19.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors16.4 GBneeds 19.3 GBRuns in the app today
  • Ovis-Image-7B

    AIDC-AI · packaged by Comfy-Org

    Ovis-Image-7B is a 7B text-to-image model built to render text well in posters, banners, logos, interface mockups and infographics.

    Runs in the app today · Images from a prompt

    Parameters
    7B
    Reads
    Text
    Makes
    Images
    Memory
    from 23.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors20.2 GBneeds 23.8 GBRuns in the app today

    Scores

  • Z-Image

    Tongyi-MAI · packaged by Comfy-Org

    Z-Image is the undistilled 6B foundation model of the Z-Image family, for making images from text at 512 to 2048 pixels per side.

    Runs in the app today · Images from a prompt

    Parameters
    6B
    Reads
    Text
    Makes
    Images
    Memory
    from 24.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors20.7 GBneeds 24.4 GBRuns in the app today
  • Z-Image-Turbo

    Tongyi-MAI · packaged by Comfy-Org

    Z-Image-Turbo is a distilled 6B text-to-image model that makes an image in 8 steps.

    Runs in the app today · Images from a prompt

    Parameters
    6B
    Reads
    Text
    Makes
    Images
    Memory
    from 24.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors20.7 GBneeds 24.4 GBRuns in the app today

    Measured by 6529 AI

    • A 1024 by 1024 image in 7 seconds · One H100, 80 GB · Nov 13, 2026
    • A 1024 by 1024 image in 4 to 5 seconds · One L40S, 48 GB · Nov 14, 2026
  • ERNIE-Image-Turbo

    Baidu · packaged by Comfy-Org

    ERNIE-Image-Turbo is Baidu's 8B text-to-image model, a distilled version of ERNIE-Image that makes an image in 8 steps.

    Runs in the app today · Images from a prompt

    Parameters
    8B
    Reads
    Text
    Makes
    Images
    Memory
    from 28.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors24.1 GBneeds 28.4 GBRuns in the app today

    Scores

  • LongCat-Image

    Meituan LongCat · packaged by Comfy-Org

    LongCat-Image is a 6B text-to-image model from Meituan's LongCat team that works in Chinese and English.

    Runs in the app today · Images from a prompt

    Parameters
    6B
    Reads
    Text
    Makes
    Images
    Memory
    from 34.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors29.5 GBneeds 34.7 GBRuns in the app today
  • Ming-Image-0.1-Design

    inclusionAI · packaged by Comfy-Org

    Ming-Image-0.1-Design is a 6B text-to-image model from inclusionAI for user interfaces, infographics, posters and other designs with a lot of text.

    Runs in the app today · Images from a prompt

    Parameters
    6B
    Reads
    Text
    Makes
    Images
    Memory
    from 58.0 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors49.3 GBneeds 58.0 GBRuns in the app today
  • Qwen-Image-2512

    Qwen · packaged by Comfy-Org

    Qwen-Image-2512 is the December update of Qwen's text-to-image model.

    Runs in the app today · Images from a prompt

    Reads
    Text
    Makes
    Images
    Memory
    from 67.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors57.7 GBneeds 67.9 GBRuns in the app today

Video generation

5 models, none in the app yet

Models that make short videos from a text prompt or a still image.

  • Wan2.2-TI2V-5B

    Wan-AI · packaged by Comfy-Org

    Wan2.2-TI2V-5B is a 5B video model that makes 720p videos at 24 frames per second from text, or from text and an image.

    Not in the app yet · The app has no route for video.

    Parameters
    5B
    Reads
    Text, images
    Makes
    Video
    Memory
    from 26.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    FP16 safetensors22.8 GBneeds 26.8 GBNot yet
  • HunyuanVideo-1.5

    Tencent · packaged by Comfy-Org

    HunyuanVideo-1.5 is Tencent's 8.3B video model that makes 480p and 720p clips from a text prompt or a still image.

    Not in the app yet · The app has no route for video.

    Parameters
    8.3B
    Reads
    Text
    Makes
    Video
    Memory
    from 42.6 GB
    License
    Tencent Hunyuan Community License Agreement (conditional)
    License terms

    The license does not apply in the European Union, the United Kingdom or South Korea, for serving or for pinning. Every copy carries the license, and a distribution carries a Notice file with the notice the license gives. Every agreement with a user must include the use restrictions and the acceptable use policy. A licensee whose products had more than 100 million monthly active users on the release date needs a license from Tencent. The license doesn't apply in EU, GB, KR.

    Versions
    VersionFilesMemoryIn the app
    FP16 safetensors36.2 GBneeds 42.6 GBNot yet
  • Wan2.2-I2V-A14B

    Wan-AI · packaged by Comfy-Org

    Wan2.2-I2V-A14B turns a still image, with an optional text prompt, into a video at 480p or 720p.

    Not in the app yet · The app has no route for video.

    Parameters
    27B total, 14B active
    Reads
    Text, images
    Makes
    Video
    Memory
    from 81.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    FP16 safetensors68.8 GBneeds 81.0 GBNot yet
  • Wan2.2-T2V-A14B

    Wan-AI · packaged by Comfy-Org

    Wan2.2-T2V-A14B makes 5-second videos at 480p or 720p from a text prompt.

    Not in the app yet · The app has no route for video.

    Parameters
    27B total, 14B active
    Reads
    Text
    Makes
    Video
    Memory
    from 81.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    FP16 safetensors68.8 GBneeds 81.0 GBNot yet
  • MiniMax H3

    MiniMax · packaged by Comfy-Org

    MiniMax H3 is a video model with a 33B transformer that makes 4 to 15 second clips with stereo sound from text, images, video or audio.

    Not in the app yet · The app has no route for video.

    Parameters
    33B
    Reads
    Text
    Makes
    Video
    Memory
    from 114.8 GB
    License
    MiniMax H3 Community License Agreement (conditional)
    License terms

    The license does not apply in the European Union, the United Kingdom, South Korea or the United States, for serving or for pinning. Every copy carries the license, and a distribution carries a NOTICE file with the notice the license gives. Every user must be bound to its use restrictions and acceptable use policy. Commercial use shows "MiniMax H3" on the user interface, and above 20 million US dollars of yearly revenue needs MiniMax's written authorization first. The license doesn't apply in EU, GB, KR, US.

    Versions
    VersionFilesMemoryIn the app
    BF16 safetensors97.6 GBneeds 114.8 GBNot yet

Embeddings and reranking

29 models, 16 in the app today

Models that turn text or images into vectors for search, and rerankers that score how well a passage answers a query.

  • all-MiniLM-L6-v2

    Sentence Transformers

    A small English sentence embedding model that turns sentences and short paragraphs into vectors of 384 numbers.

    Runs in the app today · Embeddings

    Context
    256 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF25 MBneeds 0.1 GBRuns in the app today
    F16 GGUF46 MBneeds 0.1 GBRuns in the app today
    safetensors92 MBneeds 0.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Granite-Embedding-30m-English

    IBM

    IBM's small English text embedding model of 30 million parameters, which turns text of up to 512 tokens into vectors of 384 numbers.

    Runs in the app today · Embeddings

    Parameters
    30M
    Context
    512 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF35 MBneeds 0.1 GBRuns in the app today
    BF16 GGUF63 MBneeds 0.1 GBRuns in the app today
    safetensors63 MBneeds 0.1 GBNot yet
    Its engine, PyTorch, is not in the app.
  • granite-embedding-small-english-r2

    IBM

    IBM's small English text embedding model of 47 million parameters, built on the ModernBERT design, which reads up to 8,192 tokens and returns vectors of 384 numbers.

    Runs in the app today · Embeddings

    Parameters
    47M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF52 MBneeds 0.1 GBRuns in the app today
    BF16 GGUF0.1 GBneeds 0.2 GBRuns in the app today
    safetensors0.1 GBneeds 0.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • ms-marco-MiniLM-L-6-v2

    Sentence Transformers

    A small English reranker trained on the MS MARCO passage ranking task, which reads a query and a passage together and returns a relevance score.

    Not in the app yet · The app has no route for relevance scores.

    Context
    512 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF25 MBneeds 0.1 GBNot yet
    F16 GGUF46 MBneeds 0.1 GBNot yet
    safetensors92 MBneeds 0.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • gte-modernbert-base

    Alibaba

    An English text embedding model from Alibaba's Tongyi Lab, built on ModernBERT, which reads up to 8,192 tokens and returns vectors of 768 numbers.

    Runs in the app today · Embeddings

    Parameters
    149M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.2 GBneeds 0.2 GBRuns in the app today
    F16 GGUF0.3 GBneeds 0.4 GBRuns in the app today
    safetensors0.3 GBneeds 0.4 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • gte-reranker-modernbert-base

    Alibaba

    An English reranker from Alibaba's Tongyi Lab, built on ModernBERT, which reads a query and a document of up to 8,192 tokens together and returns a relevance score.

    Not in the app yet · The app has no route for relevance scores.

    Parameters
    149M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.2 GBneeds 0.2 GBNot yet
    F16 GGUF0.3 GBneeds 0.4 GBNot yet
    safetensors0.6 GBneeds 0.8 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • ModernBERT Embed

    Nomic AI

    An English text embedding model from Nomic, trained from ModernBERT-base on the Nomic Embed data, with 768-number vectors that also work cut to 256.

    Runs in the app today · Embeddings

    Context
    8,192 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.2 GBneeds 0.2 GBRuns in the app today
    F16 GGUF0.3 GBneeds 0.4 GBRuns in the app today
    safetensors0.6 GBneeds 0.8 GBNot yet
    Its engine, PyTorch, is not in the app.
  • mxbai-rerank-xsmall-v1

    Mixedbread

    The smallest of Mixedbread's three English rerankers, which reads a query and a document together and returns a relevance score.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    512 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.2 GBneeds 0.2 GBNot yet
  • nomic-embed-text-v1.5

    Nomic AI

    Nomic's English text embedding model for search, clustering and classification, with 768-number vectors that can be shortened to as few as 64.

    Runs in the app today · Embeddings

    Context
    2,048 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.2 GBneeds 0.2 GBRuns in the app today

    Scores

    Measured by 6529 AI

    • Ran a test of embeddings. · Intel Xeon, 16 vCPUs, no GPU · Nov 14, 2026
  • bge-reranker-base

    BAAI

    A reranker for Chinese and English that reads a query and a passage together and returns a relevance score.

    Not in the app yet · The app has no route for relevance scores.

    Context
    514 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.3 GBneeds 0.4 GBNot yet
    F16 GGUF0.6 GBneeds 0.7 GBNot yet
    safetensors1.1 GBneeds 1.4 GBNot yet
    Its engine, PyTorch, is not in the app.
  • EmbeddingGemma 300M

    Google · packaged by ggml-org

    Google's first EmbeddingGemma model, a text embedding model of about 300 million parameters.

    Runs in the app today · Embeddings

    Parameters
    300M
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.4 GB
    License
    Gemma Terms of Use (conditional)
    License terms

    Every copy carries the Gemma Terms of Use. Its use restrictions, the Gemma Prohibited Use Policy, bind every user, who must be given them. Google can update the terms.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.3 GBneeds 0.4 GBRuns in the app today
  • all-mpnet-base-v2

    Sentence Transformers

    An English sentence embedding model that turns sentences and short paragraphs into vectors of 768 numbers.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    384 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.4 GBneeds 0.6 GBNot yet
  • paraphrase-multilingual-MiniLM-L12-v2

    Sentence Transformers

    A small multilingual sentence embedding model that maps text in about 50 languages into vectors of 384 numbers.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    128 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.5 GBneeds 0.6 GBNot yet
  • distiluse-base-multilingual-cased-v2

    Sentence Transformers

    A multilingual sentence embedding model that maps text in about 50 languages into vectors of 512 numbers.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    128 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.5 GBneeds 0.7 GBNot yet
  • nomic-embed-text-v2-moe

    Nomic AI

    Nomic's multilingual mixture-of-experts text embedding model for about 100 languages, with 475M parameters of which 305M work on each input.

    Runs in the app today · Embeddings

    Parameters
    475M total, 305M active
    Context
    512 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.5 GBneeds 0.7 GBRuns in the app today

    Scores

    • BEIR52.9 (the original model's score)
    • MIRACL65.8 (the original model's score)
  • BGE-M3

    BAAI · packaged by ggml-org

    A multilingual text embedding model for search in more than 100 languages, with inputs from a short sentence up to a document of 8,192 tokens.

    Runs in the app today · Embeddings

    Context
    8,192 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.8 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today
  • bge-reranker-v2-m3

    BAAI

    A multilingual reranker built on BGE-M3 that reads a query and a passage together and returns a relevance score.

    Not in the app yet · The app has no route for relevance scores.

    Context
    8,194 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBNot yet
    F16 GGUF1.2 GBneeds 1.4 GBNot yet
    safetensors2.3 GBneeds 2.7 GBNot yet
    Its engine, PyTorch, is not in the app.
  • mxbai-embed-large-v1

    Mixedbread

    Mixedbread's English sentence embedding model for search and retrieval, whose vectors can be shortened or stored in binary form to save space.

    Runs in the app today · Embeddings

    Context
    512 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    F16 GGUF0.7 GBneeds 0.8 GBRuns in the app today
  • Qwen3-Embedding-0.6B

    Qwen

    The 0.6B text embedding model in Qwen's Qwen3 Embedding series, for search, classification, clustering and code retrieval in more than 100 languages.

    Runs in the app today · Embeddings

    Parameters
    600M
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today

    Measured by 6529 AI

    • Ran a test of embeddings. · Graviton4, 16 vCPUs, no GPU · Nov 14, 2026
  • Qwen3-Reranker-0.6B

    Qwen · packaged by ggml-org

    The 0.6B reranker in Qwen's Qwen3 Embedding series, which reads a query and a document together and scores how well the document answers it.

    Not in the app yet · The app has no route for relevance scores.

    Parameters
    600M
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBNot yet

    Scores

    • MTEB-R65.8 (the original model's score)
    • CMTEB-R71.3 (the original model's score)
    • MMTEB-R66.4 (the original model's score)
    • MLDR67.3 (the original model's score)
    • MTEB-Code73.4 (the original model's score)
  • snowflake-arctic-embed-l-v2.0

    Snowflake

    Snowflake's multilingual text embedding model for retrieval in English and other languages, with inputs up to 8,192 tokens.

    Runs in the app today · Embeddings

    Parameters
    568M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today
    F16 GGUF1.2 GBneeds 1.4 GBRuns in the app today
    safetensors2.3 GBneeds 2.7 GBNot yet
    Its engine, PyTorch, is not in the app.
  • EmbeddingGemma 2

    Google · packaged by ggml-org

    A Google DeepMind embedding model that maps text, code, images, video and audio into one shared space of 768-number vectors.

    Runs in the app today · Embeddings

    Parameters
    740M
    Context
    8,192 tokens
    Reads
    Text, images, audio
    Makes
    Embeddings
    Memory
    from 1.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF1.3 GBneeds 1.6 GBRuns in the app today

    Scores

  • LaBSE

    Google · packaged by sentence-transformers

    A sentence embedding model from Google that maps text in 109 languages into one shared space of 768-number vectors.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    256 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 2.3 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.9 GBneeds 2.3 GBNot yet
  • Qwen3-VL-Embedding-2B

    Qwen

    Qwen's 2B multimodal embedding model, built on Qwen3-VL, which maps text, images, screenshots, video and mixes of them into one vector space.

    Runs in the app today · Embeddings

    Parameters
    2B
    Context
    32,768 tokens
    Reads
    Text, images, video
    Makes
    Embeddings
    Memory
    from 3.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF2.7 GBneeds 3.2 GBRuns in the app today
    BF16 GGUF4.3 GBneeds 5.1 GBRuns in the app today
    safetensors4.3 GBneeds 5.1 GBNot yet
    Its engine, PyTorch, is not in the app.
  • SigLIP 2 So400m

    Google

    Google's image and text encoder, which places pictures and short captions in one vector space and takes images at 384 by 384 pixels.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Embeddings
    Memory
    from 5.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors4.6 GBneeds 5.4 GBNot yet
  • Qwen3-Embedding-8B

    Qwen

    The 8B text embedding model in Qwen's Qwen3 Embedding series, for search, classification, clustering and code retrieval in more than 100 languages.

    Runs in the app today · Embeddings

    Parameters
    8B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Embeddings
    Memory
    from 9.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF8.1 GBneeds 9.5 GBRuns in the app today
  • Qwen3-Reranker-8B

    Qwen

    The 8B reranker in Qwen's Qwen3 Embedding series, which reads a query and a document together and scores how well the document answers it.

    Not in the app yet · The app has no route for relevance scores.

    Parameters
    8B
    Context
    32,768 tokens
    Reads
    Text
    Makes
    Relevance scores
    Memory
    from 9.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF8.1 GBneeds 9.5 GBNot yet
    BF16 GGUF15.1 GBneeds 17.9 GBNot yet
    safetensors16.4 GBneeds 19.3 GBNot yet
    Its engine, PyTorch, is not in the app.

    Scores

  • Qwen3-VL-Embedding-8B

    Qwen

    Qwen's 8B multimodal embedding model, built on Qwen3-VL, which maps text, images, screenshots, video and mixes of them into one vector space.

    Runs in the app today · Embeddings

    Parameters
    8B
    Context
    32,768 tokens
    Reads
    Text, images, video
    Makes
    Embeddings
    Memory
    from 10.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.2 GBneeds 10.9 GBRuns in the app today
    BF16 GGUF16.3 GBneeds 19.2 GBRuns in the app today
    safetensors16.3 GBneeds 19.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Qwen3-VL-Reranker-8B

    Qwen

    Qwen's 8B multimodal reranker, built on Qwen3-VL, which reads a query and a document made of text, images, screenshots or video and returns a relevance score.

    Not in the app yet · The app has no route for relevance scores.

    Parameters
    8B
    Context
    32,768 tokens
    Reads
    Text, images, video
    Makes
    Relevance scores
    Memory
    from 10.9 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF9.2 GBneeds 10.9 GBNot yet
    BF16 GGUF16.3 GBneeds 19.2 GBNot yet
    safetensors17.6 GBneeds 20.7 GBNot yet
    Its engine, PyTorch, is not in the app.

Audio and music

7 models, 2 in the app today

Models that make music or label sounds.

  • AST fine-tuned on AudioSet

    MIT

    An Audio Spectrogram Transformer fine-tuned on AudioSet that labels the sounds in a clip with the 527 AudioSet classes.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Audio
    Makes
    Labels
    Memory
    from 0.5 GB
    License
    BSD-3-Clause (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.4 GBneeds 0.5 GBNot yet
  • AST fine-tuned on Speech Commands v2

    MIT

    An Audio Spectrogram Transformer fine-tuned on Speech Commands v2 that recognizes 35 short spoken words, such as yes, no, stop and the digits zero to nine.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Audio
    Makes
    Labels
    Memory
    from 0.5 GB
    License
    BSD-3-Clause (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.3 GBneeds 0.5 GBNot yet
  • ACE-Step 1.5

    ACE-Step

    ACE-Step 1.5 makes songs from a text description and lyrics, from short loops to 10-minute pieces, with prompts in more than 50 languages.

    Runs in the app today · Embeddings; Chat

    The app serves the GGUF planner on its chat route and the GGUF text encoder on its embeddings route. The music generator runs on PyTorch, which the app doesn't run yet, so the app makes no music with them.

    Reads
    Text
    Makes
    Audio, embeddings, text
    Memory
    from 0.8 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.6 GBneeds 0.8 GBRuns in the app today
    ACE-Step's copy of Qwen3-Embedding-0.6B, the text encoder
    BF16 GGUF1.2 GBneeds 1.5 GBRuns in the app today
    ACE-Step's copy of Qwen3-Embedding-0.6B, the text encoder
    Q8_0 GGUF2.0 GBneeds 2.4 GBRuns in the app today
    the 1.7B language model that plans the song
    BF16 GGUF3.7 GBneeds 4.4 GBRuns in the app today
    the 1.7B language model that plans the song
    safetensors10.1 GBneeds 11.9 GBNot yet
    the full checkpoint: the turbo music generator, the 1.7B planner, the text encoder and the audio decoder · Its engine, PyTorch, is not in the app.
  • ACE-Step 1.5 5Hz LM 4B

    ACE-Step

    The 4B language model of ACE-Step 1.5, built on Qwen3-4B, that plans a song.

    Runs in the app today · Chat

    The app serves the GGUF files on its chat route, but this model only plans songs for ACE-Step's music generator, which runs on PyTorch and is not in the app yet. On its own it makes no music.

    Parameters
    4B
    Context
    40,960 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 5.3 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF4.5 GBneeds 5.3 GBRuns in the app today
    BF16 GGUF8.4 GBneeds 9.9 GBRuns in the app today
    safetensors8.4 GBneeds 10.0 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Magenta RealTime 2

    Google

    Google DeepMind's model that generates instrumental music continuously in real time, steered by text, audio examples and MIDI.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    2.4B
    Reads
    Text, audio
    Makes
    Audio
    Memory
    from 13.4 GB
    License
    CC-BY-4.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors11.3 GBneeds 13.4 GBNot yet
  • ACE-Step 1.5 XL SFT

    ACE-Step

    The 4B XL music generator of ACE-Step 1.5, fine-tuned for audio quality, which makes music from a text description and lyrics.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    4B
    Reads
    Text
    Makes
    Audio
    Memory
    from 23.5 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors20.0 GBneeds 23.5 GBNot yet
  • ACE-Step 1.5 XL Turbo

    ACE-Step

    The 4B XL music generator of ACE-Step 1.5, distilled to make music from a text description and lyrics in 8 steps.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    4B
    Reads
    Text
    Makes
    Audio
    Memory
    from 23.5 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors20.0 GBneeds 23.5 GBNot yet

Computer vision

21 models, none in the app yet

Models that classify images, find and outline objects, estimate depth or build 3D scenes.

  • CvT-13

    Microsoft

    CvT-13 is Microsoft's Convolutional Vision Transformer that sorts an image into one of the 1,000 ImageNet classes.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors80 MBneeds 0.1 GBNot yet
  • ResNet-18

    Microsoft

    ResNet-18 is a residual network trained on ImageNet-1k that sorts an image into one of 1,000 classes.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors47 MBneeds 0.1 GBNot yet
  • RT-DETRv2 R18

    PekingU

    RT-DETRv2 R18 is a real-time detection transformer that finds and labels objects in an image.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors81 MBneeds 0.1 GBNot yet
  • YOLOS Tiny

    hustvl

    YOLOS Tiny is a small vision transformer that finds and labels objects in an image.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors26 MBneeds 0.1 GBNot yet
  • Depth Anything V2 Small

    Depth Anything

    Depth Anything V2 Small estimates how far each part of a photo is from the camera, from a single image.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Images
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.1 GBneeds 0.2 GBNot yet
  • DETR ResNet-50

    Meta

    DETR ResNet-50 is Meta's detection transformer that finds the objects in a photo, labels them and draws a box around each.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.2 GBneeds 0.2 GBNot yet
  • DINOv2 Small

    Meta

    DINOv2 Small is a small vision transformer from Meta, trained without labels, that turns an image into features for search, classification and other tasks.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Embeddings
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors88 MBneeds 0.2 GBNot yet
  • RF-DETR Seg XXLarge

    Roboflow

    RF-DETR Seg XXLarge is Roboflow's real-time instance segmentation model, which finds objects and gives each one a mask, a box and a label.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.2 GBneeds 0.2 GBNot yet
  • SAM 2.1 Hiera Tiny

    Meta

    SAM 2.1 Hiera Tiny is a tiny-size Segment Anything 2 model from Meta that outlines objects in images and follows them through videos, starting from points or boxes.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images, video
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.2 GBneeds 0.2 GBNot yet
  • Swin Transformer Tiny

    Microsoft

    Swin Transformer Tiny is a vision transformer from Microsoft that sorts an image into one of 1,000 classes, trained on ImageNet-1k at 224 by 224 pixels.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.2 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.1 GBneeds 0.2 GBNot yet
  • SAM ViT-B

    Meta

    SAM ViT-B is the base-size Segment Anything Model from Meta, which outlines objects as masks from a clicked point or a box, or for every object in an image.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.4 GBneeds 0.5 GBNot yet
  • ViT-Base/16

    Google

    ViT-Base/16 is Google's Vision Transformer, pre-trained on 14 million ImageNet-21k images and fine-tuned on ImageNet to sort an image into one of 1,000 classes.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Labels
    Memory
    from 0.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.4 GBneeds 0.5 GBNot yet
  • Florence-2-base-ft

    Microsoft

    Florence-2-base-ft is a 0.23B Microsoft vision model that takes a short text prompt to caption an image, find objects, read text or outline regions.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    230M
    Context
    1,024 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 0.6 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.5 GBneeds 0.6 GBNot yet
  • OWL-ViT Base

    Google

    OWL-ViT Base is Google's open-vocabulary detector from 2022 that finds objects in an image from one or more short text queries.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Labels
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.6 GBneeds 0.8 GBNot yet
  • OWLv2 Base

    Google

    OWLv2 Base is Google's open-vocabulary detector that finds objects in an image from one or more short text queries.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Labels
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.6 GBneeds 0.8 GBNot yet
  • Grounding DINO Base

    IDEA Research

    Grounding DINO Base finds objects in an image from a text description, so it can detect things it was never trained to label.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Labels
    Memory
    from 1.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.9 GBneeds 1.1 GBNot yet
  • V-JEPA 2 ViT-L

    Meta

    V-JEPA 2 ViT-L is a video encoder from Meta's FAIR lab that turns a 64-frame clip or a still image into features for video classification, retrieval or a vision-language model.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Video
    Makes
    Embeddings
    Memory
    from 1.6 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.3 GBneeds 1.6 GBNot yet
  • ZoeDepth NYU-KITTI

    Intel

    ZoeDepth NYU-KITTI estimates depth in meters from a single image.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    Images
    Memory
    from 1.7 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.4 GBneeds 1.7 GBNot yet
  • Depth Anything 3 Large 1.1

    ByteDance Seed

    Depth Anything 3 Large is a 0.35B model from ByteDance Seed that estimates depth and camera positions from one or many images of a scene.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    350M
    Reads
    Images
    Makes
    Images
    Memory
    from 2.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.6 GBneeds 2.0 GBNot yet
  • MapAnything

    Meta

    MapAnything is a transformer from Meta that builds a metric 3D model of a scene from images and other inputs in one pass.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Images
    Makes
    3D objects
    Memory
    from 5.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors4.9 GBneeds 5.8 GBNot yet
  • TRELLIS.2-4B

    Microsoft

    TRELLIS.2-4B is a 4 billion parameter model from Microsoft that turns a single image into a textured 3D mesh with realistic surface materials.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    4B
    Reads
    Images
    Makes
    3D objects
    Memory
    from 19.2 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors16.2 GBneeds 19.2 GBNot yet

Robotics

3 models, 1 in the app today

Vision-language-action models that turn camera images and instructions into robot actions or plans.

  • SmolVLA base

    Hugging Face

    A compact vision-language-action model from Hugging Face's LeRobot project that turns camera views, robot state and an optional instruction into continuous robot actions.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Reads
    Text, images
    Makes
    Robot actions
    Memory
    from 1.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.9 GBneeds 1.1 GBNot yet
  • RoboBrain 2.5 4B

    BAAI

    An embodied AI model from BAAI that reads camera images and instructions and answers with 3D points, object boxes, manipulation paths and task progress for robots.

    Runs in the app today · Chat, and questions about images

    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 6.6 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF5.5 GBneeds 6.6 GBRuns in the app today
    BF16 GGUF9.7 GBneeds 11.4 GBRuns in the app today
    safetensors9.7 GBneeds 11.4 GBNot yet
    Its engine, PyTorch, is not in the app.
  • Isaac GR00T N1.7

    NVIDIA

    An open robot foundation model from NVIDIA that turns camera images, a text instruction and the robot's joint state into motor actions for humanoid manipulation.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    3B
    Reads
    Text, images
    Makes
    Robot actions
    Memory
    from 8.2 GB
    License
    NVIDIA Open Model License Agreement (conditional)
    License terms

    Every copy carries the agreement and a Notice file that reads "Licensed by NVIDIA Corporation under the NVIDIA Open Model License". A product that uses this Cosmos model shows "Built on NVIDIA Cosmos". The rights end if the model's safety guardrails are bypassed without a similar guardrail, and NVIDIA can update the agreement.

    Versions
    VersionFilesMemoryIn the app
    safetensors6.9 GBneeds 8.2 GBNot yet

Text tools and decision models

18 models, 8 in the app today

Models for one task: summaries, translation, sentiment, question answering, or a typed decision such as yes or no, a choice or a rating.

  • DistilBERT base uncased SST-2

    Hugging Face

    An English sentiment classifier from Hugging Face: DistilBERT base uncased fine-tuned on the Stanford Sentiment Treebank.

    Not in the app yet · The app has no route for labels.

    Context
    512 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF73 MBneeds 0.1 GBNot yet
    F16 GGUF0.1 GBneeds 0.2 GBNot yet
    safetensors0.3 GBneeds 0.4 GBNot yet
    Its engine, PyTorch, is not in the app.
  • T5 small

    Google

    Google's T5 at 60 million parameters, trained to treat every task as text in and text out.

    Runs in the app today · Chat

    Parameters
    60M
    Context
    512 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF65 MBneeds 0.1 GBRuns in the app today
    F16 GGUF0.1 GBneeds 0.2 GBRuns in the app today
    safetensors0.2 GBneeds 0.3 GBNot yet
    Its engine, PyTorch, is not in the app.
  • FLAN-T5 base

    Google

    A Google text-to-text model: T5 base fine-tuned on more than 1,000 tasks phrased as instructions.

    Runs in the app today · Chat

    Context
    512 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.3 GBneeds 0.4 GBRuns in the app today
    F16 GGUF0.5 GBneeds 0.6 GBRuns in the app today
    safetensors1.0 GBneeds 1.2 GBNot yet
    Its engine, PyTorch, is not in the app.
  • strands-decider-26B-A4B-gemma4-v1-2610

    Strands Agents

    A decision model from Strands Agents, built on Google's mixture-of-experts Gemma 4 26B A4B, that picks among options, answers yes or no, or rates on a scale, each with a calibrated confidence.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    25.2B total, 3.8B active
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 0.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.3 GBneeds 0.4 GBNot yet
  • TinyRoBERTa SQuAD2

    deepset

    An English question-answering model from deepset that finds the answer span in a given passage, or reports that the passage has no answer.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    514 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.4 GB
    License
    CC-BY-4.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.3 GBneeds 0.4 GBNot yet
  • RoBERTa base

    Meta

    An English text encoder from Meta, pretrained by filling in masked words.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    512 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.6 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.5 GBneeds 0.6 GBNot yet
  • Cross-Encoder NLI DeBERTa-v3 small

    Sentence Transformers

    A small English cross-encoder from Sentence Transformers, built on Microsoft's DeBERTa-v3 small and trained on SNLI and MultiNLI.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    512 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 0.7 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.6 GBneeds 0.7 GBNot yet
  • ModernBERT-base

    Answer.AI

    An English and code text encoder from Answer.AI with 149 million parameters, trained on 2 trillion tokens and reading up to 8,192 tokens.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    149M
    Context
    8,192 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.8 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.6 GBneeds 0.8 GBNot yet
  • d1-omni-600M

    Liquid AI

    A 587-million-parameter decision model from Liquid AI that answers yes/no, multiple-choice and rating questions about text, images or up to 30 seconds of English speech in one pass.

    Runs in the app today · Chat, and questions about images

    The app serves it on its chat route. Its card describes decisions read from the model's scores, which the app doesn't offer, and 6529 hasn't tested its answers through the chat route.

    Parameters
    587M
    Context
    16,384 tokens
    Reads
    Text, images, audio
    Makes
    Text
    Memory
    from 1.0 GB
    License
    LFM Open License v1.0 (conditional)
    License terms

    Every copy carries the license and its notices. Commercial use is licensed only to an entity with less than 10 million US dollars of annual revenue.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF0.9 GBneeds 1.0 GBRuns in the app today

    Scores

  • strands-decider-12B-gemma4-v1-2610

    Strands Agents

    A decision model from Strands Agents, built on Google's Gemma 4 12B, that picks among options, answers yes or no, or rates on a scale, each with a calibrated confidence.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    11.9B
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 1.0 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.8 GBneeds 1.0 GBNot yet
  • XLM-RoBERTa base

    Meta

    A multilingual text encoder from Meta, pretrained on 2.5 TB of filtered CommonCrawl text in 100 languages.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    514 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 1.4 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.1 GBneeds 1.4 GBNot yet
  • BART large CNN

    Meta

    An English summarization model from Meta: BART large fine-tuned on CNN and Daily Mail news articles paired with summaries.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    1,024 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 2.0 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.6 GBneeds 2.0 GBNot yet
  • BART large MNLI

    Meta

    BART large from Meta, trained on the MultiNLI dataset to judge whether one sentence follows from another.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Context
    1,024 tokens
    Reads
    Text
    Makes
    Labels
    Memory
    from 2.0 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors1.6 GBneeds 2.0 GBNot yet
  • d1-3B

    Liquid AI

    A 3.1-billion-parameter decision model from Liquid AI that takes text, JSON or images plus a set of questions and returns calibrated yes/no answers, picks among named options or ratings in one pass.

    Runs in the app today · Chat, and questions about images

    The app serves it on its chat route. Its card describes decisions read from the model's scores, which the app doesn't offer, and 6529 hasn't tested its answers through the chat route.

    Parameters
    3.1B
    Context
    32,768 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 4.4 GB
    License
    LFM Open License v1.0 (conditional)
    License terms

    Every copy carries the license and its notices. Commercial use is licensed only to an entity with less than 10 million US dollars of annual revenue.

    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF3.7 GBneeds 4.4 GBRuns in the app today

    Scores

    • Decision Index 0.2.148.6 (the original model's score)
    • MMBench84.9 (the original model's score)
    • POPE88.5 (the original model's score)
    • SQuAD 2.085.3 (the original model's score)
    • BoolQ86.7 (the original model's score)
  • Kev-9B

    jaredpalmer · packaged by ggml-org

    Kev-9B is a decision model built on Qwen3.5-9B-Base with a small adapter and scoring head.

    Runs in the app today · Chat

    The app serves it on its chat route. Its card describes a scoring interface of its own, which the app doesn't offer, and 6529 hasn't tested its answers through the chat route.

    Parameters
    9B
    Context
    65,536 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 7.5 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF6.4 GBneeds 7.5 GBRuns in the app today

    Scores

  • Clef-Flash

    Cloudflare · packaged by ggml-org

    Clef-Flash is the 9-billion-parameter version of Cloudflare's Clef decision model, post-trained from Qwen3.5-9B.

    Runs in the app today · Chat, and questions about images

    The app serves it on its chat route. Its card describes a scoring interface of its own, which the app doesn't offer, and 6529 hasn't tested its answers through the chat route.

    Parameters
    9B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 9.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF7.7 GBneeds 9.1 GBRuns in the app today

    Scores

  • Hy-MT2-7B

    Tencent

    A 7-billion-parameter translation model from Tencent that translates among 33 languages.

    Runs in the app today · Chat

    Parameters
    7B
    Context
    262,144 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 9.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q8_0 GGUF8.0 GBneeds 9.4 GBRuns in the app today
  • Clef

    Cloudflare · packaged by ggml-org

    Clef is a 27-billion-parameter decision model from Cloudflare, post-trained from Qwen3.8-27B, that reads text, JSON, images or video.

    Runs in the app today · Chat, and questions about images

    The app serves it on its chat route. Its card describes a scoring interface of its own, which the app doesn't offer, and 6529 hasn't tested its answers through the chat route.

    Parameters
    27B
    Context
    262,144 tokens
    Reads
    Text, images
    Makes
    Text
    Memory
    from 24.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF20.5 GBneeds 24.1 GBRuns in the app today

    Scores

Science and forecasting

3 models, 1 in the app today

Models for DNA, protein sequences and time series.

  • ESM-2 8M

    Meta

    A protein language model from Meta with 8 million parameters and 6 layers, trained to fill in masked amino acids.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    8M
    Context
    1,026 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 0.1 GB
    License
    MIT (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors31 MBneeds 0.1 GBNot yet
  • TimesFM 2.5 200M

    Google

    A pretrained time-series forecasting model from Google that predicts the next values of a numeric series from its history, giving point and quantile forecasts.

    Not in the app yet · Its engine, PyTorch, is not in the app.

    Parameters
    200M
    Reads
    Makes
    Memory
    from 1.1 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    safetensors0.9 GBneeds 1.1 GBNot yet
  • Carbon-X

    Hugging Face

    Carbon-X is a 14.8-billion-parameter chat model from Hugging Face, trained from Qwen3-14B on DNA sequences and English text.

    Runs in the app today · Chat

    Parameters
    14.8B
    Context
    40,960 tokens
    Reads
    Text
    Makes
    Text
    Memory
    from 11.4 GB
    License
    Apache-2.0 (permissive)
    Versions
    VersionFilesMemoryIn the app
    Q4_K_M GGUF9.6 GBneeds 11.4 GBRuns in the app today

    Scores