Highlights
Comprehensive Multi-Format Support for Mainstream Models
This release further refines the support for the following models across different inference engines and model formats, including vLLM, Transformers, Quantized models, and GGUF models:
GLM-5
Kimi-K2.5
MiniMax-M2.5
Qwen3.5
Notably, we have focused on fixing and improving the compatibility of Qwen3.5 across multiple inference engines, ensuring more stable deployment in various runtime environments.
Community Edition Updates
Installation
- ❯ pip install 'xinference==2.3.0'
- ❯ Docker: Pull the latest image or update via pip inside the container.
New Features
- vLLM engine support for Qwen3.5.
- Added seed and repetition_penalty parameters for precision testing and inference control.
Enhancements
- Transformers engine support for Qwen3.5.
- Continuously updated model JSON configurations to improve compatibility.
Bug Fixes
- Fixed the parameter conflict issue (
rpc_rank) inWorkerWrapperBase.__init__(). - Fixed the check issue for
qwen3-vl-embeddingunder vLLM. - Fixed multi-GPU chat inference issues.
- Fixed the startup issue of Qwen3.5 under vLLM caused by missing
generation_config.json. - Fixed the UI form default value initialization issue.
Documentation
- Added v2.2.0 release notes.
- Corrected various documentation details.
Enterprise Edition
🎉 Xinference Enterprise Edition 1.0 Officially Released
The first official release of Enterprise Edition 1.0 marks a significant milestone, signifying that Xinference's enterprise-grade capabilities have reached a stable and mature stage.