Test release0.1.0-rc.2
Known issues
This is an untested test release: expect problems, and don't rely on it for anything that matters. Below: the bugs found so far, then what hasn't been tested.
Updated November 17, 2026
Found a problem? Report it
Known bugs
Found in real use, with its fix.
A model about as big as the computer's memory can take many minutes to load from a slow disk.
Fix: Keep the models on a fast local disk, or pick a smaller model: from a local NVMe disk the same 15.6 GB model loaded in 70 seconds.
Why: The engine maps the model's files into memory, and when the memory can't hold them all it reads the same parts from the disk again and again.
Fixed in 0.1.0-rc.2
- On a fresh Windows, Fetch and serve no longer fails with "error sending request for url".
- The Windows installer picks the GPU engines only when the computer has a graphics card.
- The sign-in to the page lasts through a restart of the service.
- A long chat title no longer hides Rename and Delete.
- A model's attempt to call a tool shows as a plain note in the chat.
- Chat models on a Mac start with a 16K context.
On an earlier release? Download the newest and install it.
Not tested yet
What this release hasn't run on, by system.
Mac
Tested only on a cloud Mac (M2, macOS 15.8.2).
Not tested on macOS before 15.8; it is built for 13.3 and later.
Windows
Tested only on Windows Server 2025, not on Windows 10 or 11.
Not tested with a GPU.
Workaround: If models don't start, reinstall and choose On the CPU only. If the uninstall leaves the config behind, delete it first:
File to deleteC:\Program Files\6529 AI\6529-ai.tomlOpening the installer by double-click in File Explorer, and SmartScreen's warning, haven't been tested; it was opened from PowerShell.
Linux
Tested only on Ubuntu 22.04 and 24.04.
Workaround: On another distribution, the
.tar.gzinstalls the same files.The CUDA packages are tested only on an NVIDIA L4, not on other NVIDIA GPUs.
On every system
The page on a phone hasn't been tested: it answers on 127.0.0.1 only, and the phone-width pictures on Use it are a desktop browser made narrow.
Record, on the Speech screen, hasn't been tested.
If the page doesn't open within 60 seconds of installing, the service didn't start.
Workaround: Its log says why:
Mac, log file~/Library/Logs/6529-ai.logWindows, log file%LOCALAPPDATA%\6529-ai\logs\6529-ai.logLinux, Terminaljournalctl -u 6529-ai
Found a problem?
Say which system you're on, what you did, and what happened, and add a screenshot and the log.
