CommunityAI alpha — community inference without model downloads
Pre-release
Community inference now runs entirely on contributing peers. Your app sends messages and receives answers without downloading community model weights or running its input/output stages locally.
- The small local Qwen fallback remains available when the community cannot answer. The explicit local-only choice is preserved.
- A 100% sharing limit now makes the full GPU memory capacity available for sharing. No memory is permanently reserved for an unused fallback.
- Model health uses the shared discovery status, removing the separate local tensor router that could show zero available blocks while the public mesh was complete.
- The desktop accepts the new zero-byte community download status.
Existing September 9 updater installations download this release automatically and offer Restart to update. Click the update status in the sidebar to check immediately. Earlier installations need one manual installer upgrade. Settings and model caches are preserved.
| Platform | Small download | Complete offline installer |
|---|---|---|
| Windows | Online setup | Offline setup |
| Ubuntu/Debian | Online installer | Offline .deb |
The online installers fetch and verify the complete runtime. Windows publisher signing remains deferred for alpha. Checksums.
Two real public-mesh requests completed with an empty consumer cache and all artifact downloads forbidden. The temporary CPU mesh took about 107 seconds for a short completion and 158 seconds for a short chat; this is functional evidence, not a latency target. Text-serving peers currently require operator setup through the source CLI (drift text-peer); automatic placement of that role on desktop contributors remains future work.
Fresh Windows and Linux packages use source d7f4333cc8ff74ce060076c1d6effb67fb4c6d2c and the normal packaging checks. Large hosted installers are checked through object size, SHA-256 metadata and public start/end samples; small public artifacts receive full hash checks. This release does not claim a new full installed desktop or GPU test campaign.