Xinference 3.3

✨ New Features

• ▸ jina-reranker-m0 rerank model
• ▸ WebUI shows downloaded file sizes
• ▸ qwen3.8-27B and complete Qwen3.8 support
• ▸ WebUI IndexTTS emotion controls
• ▸ Multi-engine audio with MLX models
• ▸ minimax music3 model
• ▸ MiniMax-H3 Lightning LoRA
• ▸ FireRedTTS3 audio support
• ▸ DeepSeek V4 Flash 0731 registration
• ▸ Ideogram4, SenseNova U1.5, HiDream-O1, GLM-Image & Krea 2 image models
• ▸ Token-aware virtual model routing
• ▸ Tokenizer asset registry with a bundled DeepSeek V4 asset
• ▸ Media seed controls and dimension swap
• ▸ Token Router management UI
• ▸ Agent-based Router orchestration
• ▸ Ornith-1.5-35B-A3B and Ornith-1.5-397B
• ▸ Token Router observability
• ▸ Router Agent and cluster management UI
• ▸ Unified Anthropic Messages protocol adapter, routable via virtual models
• ▸ glm-5.2 and Kimi-K3 model support
• ▸ Secure per-launch virtualenv find-links, incl. the Jina flash-attn wheel
• ▸ Per-launch virtualenv find-links in the UI
• ▸ Expose virtual model running details
• ▸ ACE-Step 1.5 music generation
• ▸ World model generation support
• ▸ Replica config in the launch CLI and Web UI
• ▸ Configurable dynamic replica scale-up
• ▸ Persistent system settings management
• ▸ NaviDC-OCR with Transformers and vLLM
• ▸ WeMM-Embedding support
• ▸ Breeze-TTS-2 support
• ▸ Show full paths on hover

🛠 Enhancements

• ▸ Updated model configs (rerank / video)
• ▸ Adaptive image previews in the UI
• ▸ VoiceDesign ability exposed on Qwen3-TTS-Voice-Design
• ▸ Multi-engine video support with MLX
• ▸ Configurable speech and music output formats
• ▸ Base64 video input in vLLM chat

🐞 Bug Fixes

• ▸ Correct IndexTTS-2.5 runtime dependencies
• ▸ Return playable speech responses
• ▸ Pin DeepDoc transformers to >=4.51,<5 for OCR launches
• ▸ Auto-size launch model dropdowns and include the scrollbar in width
• ▸ Guard missing/invalid usage in non-stream chat metrics
• ▸ Calculate download progress per file
• ▸ Validate LLM metadata before packaging
• ▸ Refine autostart transition handling and reset retry attempts
• ▸ Wait reliably for metrics exporter startup
• ▸ Make worker supervisor reconnection generation-safe
• ▸ Harden the batch processor against cancellation and unexpected exits
• ▸ Normalize distributed worker count
• ▸ Preserve DeepSeek V4 tool argument types
• ▸ vLLM: handle engine death during async iteration; block size 256 for DeepSeek V4
• ▸ Map Qwen3.8 top-level reasoning effort
• ▸ Pin Qwen Layered dependencies
• ▸ Fully remove downloaded models from cache
• ▸ Reschedule autostart when the last replica or whole worker dies
• ▸ Honor configured sources in post-install hooks
• ▸ Clarify Token Router identifiers and harden running-model and detail views
• ▸ Separate runtime and backend credentials; stabilize managed runtime identity
• ▸ Support Hugging Face cache roots in auto-fill
• ▸ Stabilize MLX runtime dependencies and locks
• ▸ Declare the Token Router installation extra
• ▸ Skip cache status update when the cache tracker ref is None
• ▸ Preserve available model download hubs
• ▸ Classify OvisOCR2 as OCR
• ▸ Stabilize HY-WorldPlay runtime and progress reporting
• ▸ Restore GPU and aarch64 Docker builds
• ▸ Repair pypiserver platform dependencies

📚 Documentation

• ▸ Add v3.2.0 to the release notes index and locale catalogs
• ▸ CI: skip GPU tests for video and flexible model changes
• ▸ Router: remove legacy Agent compatibility
• ▸ Refresh the built-in model catalog and skip draft CI