🚀 New Features
- Recommended models are refreshed, Qwen3.6-27B and Qwen3.6-35B-A3B now show up across the 24, 32, 48 and 64 GB tiers
- The Vulkan GPU build is offered to far more Linux machines now, any GPU with 2 GB or more counts
- New "Fit to device memory" switch lets llama.cpp pick the context size for you, it is on by default and the UI shows what the engine settled on
- Model scanning now picks up GPT4All, Jan, Msty and llama.cpp cache folders, and respects OLLAMA_MODELS, HF_HOME and the rest
- Sending a message starts the right model on its own, last used first, then cloud, then whatever local model you have, and only asks you when the choice is genuinely unclear
🔧 Improvements & Fixes
- A model that runs out of memory no longer just fails, the app halves the context, then falls back to CPU, and tells you what it did
- KV cache size is estimated properly for every cache type, so fit checks and compatibility badges match reality
- Your context length stops getting overwritten by the provider, and a few dead engine controls are gone
- MLX keeps your context size across drafter reloads and no longer trips over numeric settings
- Onboarding only counts models that can actually answer, and MLX counts as a local provider now
- A bunch of smaller fixes across the engine and setup
🙏 Contributors
Thanks to @Vect0rM for their contributions to this release!