yue2-v0.1.6 v0.76.0
Gen-Verse/MMaDASep 9, 2026 → Sep 10, 20261 day between releases
What's New
Switch from YuE2 music generation to Mesh-LLM LLM serving infrastructure.
Breaking Changes
- Stop persisting KV cache state to disk
- Remove stale disk cache defaults
- Treat max_tokens as a ceiling, not a context reservation
- Consolidate model=auto and model=mesh into one automatic-routing directive
- Replace flat prefix caches with unified radix
New Features
- Durable KV prefix cache that survives eviction and restart
- Skippy speculative decoding and Qwen3.8 Flash-Next support
- Stream tool calls as they are generated
- Mode-aware management health endpoint
- OpenAI exchange lifecycle hooks for plugins
Improvements
- Significant improvements to model serving, routing, and observability
- Iteration-level scheduler for concurrent staged serving
- Route repeated prompts to workers with verified radix cache reuse
- Reuse warm prefixes across agent sessions
- Stream tool calls as they are generated instead of after the turn ends
Upgrade Notes
You must stop persisting KV cache state to disk and remove stale disk cache defaults. Treat max_tokens strictly as a ceiling and consolidate model directives into a single automatic-routing setting.
The Two Releases
YuE2 · frontier music generation with symbolic planning
Sep 9, 2026· 1 month ago
Introduces frontier music generation capabilities with symbolic planning, enabling high-quality song creation from lyrics and zero-shot covers from transcribed melodies.
Full release notesv0.76.0
Sep 10, 2026· 1 month ago
A major feature release introducing durable KV caching, speculative decoding (Skippy), and significant improvements to model serving, routing, and observability. It also updates the configuration system and request handling logic.
Full release notes