Skip to content

vMLX v1.3.25

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 03 Apr 01:43
· 3830 commits to main since this release
v1.3.25
02c3058

What's New

Gemma 4 Full Support

  • Gemma 4 reasoning parser — properly handles <|channel>thought...<channel|> protocol (fixes raw token leaking in chat)
  • Gemma 4 tool parser — supports native <|tool_call>call:name{args}<tool_call|> format
  • Gemma 4 stop tokens — <turn|> correctly added as stop token alongside <eos>
  • Works with all thinking modes (Auto/On/Off)

Engine Fixes

  • Extra eos_tokens from model config now applied to both LLM and MLLM schedulers
  • <|tool_call> added to tool call marker detection for streaming
  • GitHub issues #43-#46 fixes
  • Deep API/gateway audit (11 issues fixed)
  • Ollama auth fix, tool auto-detect from request
  • --default-enable-thinking CLI flag
  • TQ-native disk cache improvements

Full Changelog: v1.3.24...v1.3.25

Assets 2
Loading