Xinference v2.5.0 Released

Empowering your AI infrastructure with new models, enhanced stability, and powerful web UI features.

✨ Highlights

New Model Support

  • Qwen3-TTS Series (0.6B / 1.7B)
  • Gemma-4

Web UI Config Management

  • Save multiple deployment configs locally
  • Easy switching between parameter sets

Auto Recovery

  • Supervisor auto-reconnects after restart
  • Restores model replica states automatically

Install Get Started

pip install 'xinference==2.5.0'

Or pull the latest Docker image and update via pip inside the container.

🆕 New Models & Features

Models

  • Qwen3-TTS-12Hz-0.6B / 1.7B (Base / CustomVoice / VoiceDesign)
  • Gemma-4

Features

  • sglang support for Qwen3.5
  • Added Qwen3-TTS audio model support
  • WebUI local config saving
  • Supervisor auto-recovery

🛠 Enhancements

  • Automatic detection of PyTorch CUDA version
  • Lightweight worker liveness detection
  • Optimized deployment progress & status display
  • Continuous model JSON config updates

🐞 Bug Fixes

Fixed auth credential comparison security issue (CWE-208)
Fixed Qwen3 reranker accuracy under vLLM
Fixed Qwen3.5 streaming API tag error
Fixed Jinja template `{% break %}` parsing (Kimi K2.5)
Fixed qwen3-vl embedding issue under vLLM

🏢 Enterprise Edition

Version 1.0.2 Released

Extensive stability fixes and optimizations included in this release.

🎉 Community

Maintainer Promotion

Congratulations to @ZhikaiGuo960110 for becoming a Xinference Maintainer!

📚 Documentation updated with v2.4.0 release notes.