🚀 New Features
New models on llama.cpp TurboQuant
- DeepSeek V4 Flash 0731 — 284B MoE, 13B active, 1M context, re-post-trained for coding and agents
- Kimi K3 — 2.8T MoE, 104B active, 1M context. Native vision included
- Ling-3.0-flash — 124B hybrid-reasoning MoE, only 5.1B active per token
- LFM2.5-2.6B — fully on-device agentic model: plans, calls tools, 128K context
🔧 Improvements & Fixes
- AMD on Linux x64 now covers RDNA2–RDNA4 and CDNA on llama.cpp TurboQuant
- Synced with the latest upstream llama.cpp TurboQuant
- Engines now update on their own — the app detects your hardware and pulls the right TurboQuant build, so new engine releases land without waiting for an app update
- Autostart setting now actually sticks
- Downloads no longer fail when the pre-check does
- Clearer model-load errors, and recoverable ones now retry
- Windows: models in folders with non-Latin names now load
- MLX models with legacy or misnamed shards now load
- Favorites no longer list the same model twice
- Proper icon for the Codex integration
- Clearer wording around engine update checks
- Smaller UI and stability cleanups
🙏 Contributors
Thanks to @Vect0rM, @worthant, @yl-atomic for their contributions to this release!