Model download recovery and inference compatibility
Improve model download recovery, adjust the default reasoning budget to 8K tokens, and update local inference compatibility with UTF-8 output guarding.
More reliable model download recovery
Pro/Lite downloads preserve progress and retry after temporary network failures. Restarting the app can continue unfinished downloads. Setup and Settings share download status, with clearer controls for pausing, cancelling, and restarting individual files. Manually paused tasks remain paused.
Reasoning budget and inference compatibility
The default reasoning budget is now 8K tokens. The local inference engine adds GPU compatibility fallbacks across Apple chips and UTF-8 output guarding to reduce output problems caused by invalid byte sequences.
Upgrade notes
This release supports Apple Silicon Macs running macOS 14 or later. Fully downloaded MLX Pro/Lite models remain usable without another download. Existing GGUF users need to download the corresponding MLX models; original model files, conversations, and settings are retained.