Xinference v2.12.0
Model autostart lands, security & permissions get a major upgrade, and 4 new models arrive
📅 Released: July 6, 2026
📦 Installation
pip install 'xinference==2.12.0'
Pull the latest image or update via pip
✅ Highlights
Model Autostart
A core capability: models can automatically resume their running state after system startup or recovery, cutting manual intervention.
Major Security & Permissions Upgrade
Stronger remote-code execution controls, upgraded API Key / user / permission system, and clearer resource isolation.
Expanded Model Support
VibeThinker 1.5B/3B, the Nex-N2 series, Ornith-1.0-35B, and Unlimited-OCR are now available.
Web UI & Interaction Upgrades
New API Key / user management pages, plus a "Try to API" quick-test option for running models.
Embedding / Rerank Enhancements
Richer capability panels and clearer model capability management.
Monitoring & Observability
Grafana dashboards split into 6 sub-panels; Prometheus metrics unified with a new replica_index metric.
🆕 New Models
VibeThinker 1.5B / 3B
Supports both Transformers and vLLM inference engines.
Nex-N2 (mini / Pro / Pro-FP8)
The Nex-N2 family launches, covering lightweight to high-performance scenarios.
Ornith-1.0-35B
A large-parameter model that extends local inference capability.
Unlimited-OCR
A new OCR model, extending document recognition and parsing scenarios.
✨ New Features
Model Autostart
Models can automatically resume their running state after system startup or recovery.
Stronger Remote Code Execution Security
Hardened security checks for remote code execution.
API Key / User Management Pages
New pages for managing API keys and users.
Embedding / Rerank Panel Enhancements
Clearer capability display and improved model capability management.
Running Model Quick API Test
Running models now support "Try to API" for quick verification without leaving the UI.
Unified Worker / Supervisor / Replica Metrics
Consistent metric naming across roles for easier monitoring and alerting.
📊 Monitoring Enhancements
Grafana Dashboard Split into 6 Sub-Views
Panels broken down by dimension for faster troubleshooting.
Enhanced Prometheus Metrics
Unified worker / supervisor / replica metrics, plus a new replica_index metric.
🛠 Improvements
- Log rotation now supports daily + size-based policies
- Multi-process safe log writing
- Enhanced structured logging
- Multi-select Worker IP filter component
- Improved UI menu / permission system
- Streamlined model loading and registration
🐞 Bug Fixes
- Fixed embedding / rerank model dependency issues
- Fixed vLLM / llama.cpp / tool call compatibility issues
- Fixed abnormal GPU / worker status issues
- Fixed OAuth / permission escalation issues
- Fixed metric loss caused by cache initialization failures
📚 Documentation
- Added multi-language documentation support
- Added a security policy
- Updated internationalization and documentation structure
🏢 Enterprise Edition Updates
- 🔧 Continued stability and reliability improvements