v4.3.5
📦 localaiView on GitHub →
✨ 1 features🐛 5 fixes🔧 1 symbols
Summary
This release introduces the ability to honor per-request reasoning effort for chat completions and includes several bug fixes related to tool-call streaming and application shutdown handling.
✨ New Features
- Honor per-request reasoning_effort on chat completions.
🐛 Bug Fixes
- Fixed tool-call JSON leaking into content when using stream+tools on tokenizer-template models.
- Stopped double-emission of streaming tool-calls when the autoparser is active.
- Ensured backend processes stop synchronously on shutdown.
- Validated auto-detected XML tool-call names, adding robustness for glm-4.5/Hermes.
- Improved model crash detection by tracking intentional stops, preventing misreading clean shutdowns as crashes.