⚛️ Meet Atomic Chat Core
Atomic Chat now runs on its own open-source server, Atomic Chat Core. It ships inside the app and updates with it. The core runs local models on llama.cpp, MLX and TurboQuant. It also handles downloads, image and video generation, the Local API Server and cloud providers. It is built to work with several inference engines, so Atomic Chat can support more engines and model types over time.
What changes for you:
- Model loading, downloads and the API server now go through one place
- The Local API Server comes back on its own after a restart with the model it was serving, tells you when it had to use a different port, and now serves image and video models at
/v1/images/generationsand/v1/videos - Settings → General shows which core version you're running
- Logs from the app and the core land in one timeline and export as one file you can attach to a bug report
- Your chats, models and settings carry over as they are
🚀 New Features
- Generate short videos locally with LTX-2.3 or Wan 2.2, see how long a clip will take and whether it fits in memory before it starts, and follow a progress bar with the time left in the preview, with a stop button. Image to video is coming next
- The media engine for images and video installs in one click, and the studio opens as soon as it's ready
- The Models Hub has Text, Images and Video tabs, and image and video model cards show their Hugging Face README, download size and whether they fit your memory
- Speed & Stability of downloads increased. Big downloads come in over several connections at once, reconnect on their own when a connection goes quiet, and resume where they stopped
- Desktop notifications tell you when a download, an image or a video is ready
- Eden AI joins the list of cloud providers
- Remote & LAN moved to the API page, next to the server it controls
🔧 Improvements & Fixes
- Image and video generation falls back to offloading when the GPU runs out of memory instead of failing, and Auto keeps models on a discrete GPU when you have one
- Image and video settings no longer reset when you start a model, Generate starts an installed model for you, and time estimates stay accurate on long clips
- On Macs, Fit to device memory leaves half of RAM to the rest of the system, so a small model no longer eats most of your memory for its context
- Multi-token prediction works on more model families, for faster replies on models that have it built in
- Local replies no longer lose the end of a tool call or break emoji and non-Latin characters mid-stream
- After a model fails to load, the composer asks you to pick a model instead of sending to one that isn't running
- Windows no longer goes "Not Responding" while the app reads your hardware (#326)
- Local models are unloaded before an update installs, so updates no longer hang
- Onboarding no longer gets stuck on the Welcome screen after you start a download, and models you downloaded during setup show up as installed in the Hub
- Hub search has a clear button
- Assistants still using the old Jan default prompt get the Atomic Chat one, custom prompts stay untouched
- Factory reset also clears the core's data and saved keys, and keeps your downloaded engines
- The Apple on-device provider is hidden for now, it could report itself ready and then fail to answer
- Concurrent Mode is gone from llama.cpp settings
- 50+ smaller stability and UI fixes across downloads, the Hub, the model picker, onboarding and the media pages
🙏 Contributors
Thanks to @Vect0rM, @danyurkin, @Ooooze, @xDenside and @AlexFromAtomic for their contributions to this release!