🚀 Xinference v2.1.0 Release Notes

✅ Highlights

• 🧠 GLM-4.7 / GLM-4.7-Flash Support
  Added full support for GLM-4.7 and GLM-4.7-Flash, further expanding the GLM model ecosystem.
• 🎤 Qwen3-ASR Series Launched
  New additions:
  • Qwen3-ASR-0.6B
  • Qwen3-ASR-1.7B
  Fully supports Qwen3-ASR speech recognition models, covering both lightweight and high-performance scenarios.
• 🖼️ FLUX.2-Klein Series Support
  New additions:
  • FLUX.2-Klein-4B
  • FLUX.2-Klein-9B
  Enhanced image generation and editing capabilities, continuously improving FLUX ecosystem support.
• 🔁 MinerU2.5-2509-1.2B Adjustment
  Updated and adjusted the MinerU2.5-2509-1.2B model, optimizing model configuration and adaptation processes.

🌐 Community Edition Updates

📦 Installation

• pip: pip install 'xinference==2.1.0'
• Docker: Pull the latest image, or update via pip inside the container.

🆕 New Model Support

• GLM-4.7
• GLM-4.7-Flash
• Qwen3-ASR-0.6B / 1.7B
• FLUX.2-Klein-4B / 9B

🛠 Enhancements

• Updated DeepSeek-V3.2 / DeepSeek-V3.2-Exp model configurations.
• Optimized image build dependencies (constrained setuptools < 82).
• Refactored API layer structure:
  • Extracted Pydantic request Schemas.
  • Modularized route registration for clearer code structure.

🐞 Bug Fixes

• Fixed vLLM embedding model error.
• Fixed vLLM reranker scoring anomaly.
• Fixed vLLM reranker GPU release issue.
• Compatible with vLLM async tokenizer logic.
• Fixed CI issues related to setuptools.

📚 Documentation

• Added v2.0.0 release notes.

🏢 Enterprise Edition Updates

• 🔧 Stability Enhancements
  Includes multiple underlying optimizations and bug fixes to improve overall runtime stability and enterprise deployment reliability.

---

Welcome to upgrade and experience Xinference v2.1.0 🚀