Starstruk Logo MarkStarstruk

v0.1.806-beta v1.0.0

high risk

unslothai/unslothSep 2, 2026 → Sep 6, 20264 days between releases

What's New

Unifies speech-to-speech engine with new serve/talk/local commands

Breaking Changes

  • Commands `--mode realtime` and `--mode local` deprecated/removed in favor of explicit commands
  • `--mac-optimal-settings` removed from main command, use `local` or `serve` instead
  • Redundant extras `facebook-mms`, `language-detection`, `websocket` removed
  • Default hosted LLM changed to `gpt-5.6-terra`

New Features

  • Qwen3-ASR speech recognition
  • OpenAI-compatible STT and TTS endpoints
  • Stateful OpenAI Realtime transcription
  • Experimental vLLM Realtime transcription
  • Optional OmniVoice and Supertonic TTS

Improvements

  • Improved ordered text, audio, and tool-call output handling
  • Updated Apple Silicon setup with optimized MLX dependencies
  • Clearer setup and diagnostics

Bug Fixes

  • Smoother model loading across local servers and connected providers
  • Safer chat edits that preserve tool cards and reply details
  • Improved conversation reliability with interrupted responses

Upgrade Notes

Update your CLI commands to use `serve`, `talk`, or `local` instead of `--mode realtime` or `--mode local`. Remove deprecated extras like `facebook-mms` and `websocket` from your installation commands. Note that the default hosted LLM has changed to `gpt-5.6-terra`.

The Two Releases

v0.1.806-beta

2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP

Sep 2, 2026· 2 months ago

Significantly boosts inference speed for Qwen and GLM models using MTP, introduces new audio and video generation APIs, and enhances MLX support.

Full release notes
v1.0.0

v1.0.0

Sep 6, 2026· 2 months ago

The first major release of speech-to-speech unifies the Realtime engine with three commands (`serve`, `talk`, `local`), adds extensive speech backends like Qwen3 and OpenAI-compatible endpoints, and improves conversation reliability.

Full release notes