Release Notes · July 2026

Xinference 3.1

Enterprise single sign-on, a multi-dashboard monitoring center, speculative decoding, and new inference engines — a focused upgrade on the 3.0 foundation.

🔐 SSO / OIDC 📊 Monitor Center ⚡ Speculative Decoding 🧩 New Engines

What's New

Highlights

🔐

SSO / OIDC Login

A dedicated SSO/OIDC login button on the login page brings seamless enterprise single sign-on to every deployment.

  • • One-click OIDC login button
  • • Generic SSO branding, not vendor-locked
  • • New permission scopes surfaced in the UI
📊

Monitoring & Audit

A multi-dashboard monitor center with an admin config dialog, plus a richer, auto-refreshing audit center.

  • • Multi-dashboard monitor center
  • • Audit center auto-refresh
  • • API key name & model ID columns
⚡

Speculative Decoding

Speculative decoding with paired drafter checkpoints accelerates generation without sacrificing output quality.

  • • Paired drafter checkpoints
  • • Faster token throughput
  • • vLLM logprobs / top_logprobs support
🧩

New Engines & Models

New sglang image and vLLM ASR engines, batched Qwen3-ASR transcription, and day-one MiniMax-M3 support.

  • • sglang image & vLLM ASR engines
  • • Batch Qwen3-ASR transcriptions
  • • MiniMax-M3 model support

Open Source

Community Edition

Installation

pip

pip install 'xinference==3.1.0'

Docker

docker pull xprobe/xinference:latest

✨ New Features

  • ▸ SSO/OIDC login button on the login page
  • ▸ Multi-select model launch across workers
  • ▸ Batched Qwen3-ASR transcriptions
  • ▸ Multi-dashboard monitor center with admin config
  • ▸ Auto-refresh for the audit center
  • ▸ API key name & model ID columns in audit center
  • ▸ Multi-node selection in the log center
  • ▸ Auto-reload shell routes on export directory change
  • ▸ vLLM logprobs / top_logprobs in completion responses
  • ▸ Speculative decoding with paired drafter checkpoints
  • ▸ sglang image & vLLM ASR engines
  • ▸ Missing permission scopes (models:register, logs:list, monitor:view) in UI
  • ▸ MiniMax-M3 model support

🛠 Enhancements

  • ▸ Updated model configs (embedding / rerank / audio / llm / image)
  • ▸ Replica launch status shown in a responsive grid
  • ▸ xoscar 0.9.7 graceful #system_*# fallback
  • ▸ aarch64 image built from the unified slim Dockerfile
  • ▸ Improved Docker build caching
  • ▸ peft>=0.18.0 for transformers 5.x compatibility

🐞 Bug Fixes

  • ▸ sentence_transformers load failures from torch-ecosystem mismatches
  • ▸ Slow HTTP DoS (Slowloris) mitigation via request deadline
  • ▸ Pin torch to system version to prevent torchvision ABI crash
  • ▸ Stale model state after a failed async engine load
  • ▸ MLX VLM generation kept on a dedicated thread
  • ▸ Disambiguate replica model UIDs with a -rep{n} suffix
  • ▸ Re-normalize per-item embedding index after batch split
  • ▸ Stabilize Jina v5 & Qwen3 reranker deployment
  • ▸ Worker reconnection after supervisor restart
  • ▸ Skip redundant virtualenv setup when replicas share a venv
  • ▸ TTS on Windows: don't reopen a NamedTemporaryFile

📚 Documentation

  • ▸ Fix v3.0.0 Chinese release links
  • ▸ Update release notes links to xinference.co