What's New
GLM-5 / DeepSeek-V3 CacheList support
- Fix SSD cache serialization for models with zero-dimension tensors (DSA indexer)
- Enable per-block prefix cache restore for CacheList models (previously always rejected partial matches)
Tool calling improvements (#7)
- Fix tool calling by preserving native format in conversation history
- Add MiniMax/namespaced tool call format parser
- Client disconnect detection during prefill in streaming
Admin dashboard
- Engine versions display and model name copy buttons
macOS app
- Welcome screen step card spacing fix
- GitHub link button in About dialog
- Update check button opens GitHub releases page