v1.9.0 v1.10.0
cheaper-inference/OmniRouteMar 1, 2026 → Mar 5, 20263 days between releases
What's New
Adds VibeVoice training UI, Qwen3.5 support, and Python 3.11 default
Breaking Changes
- Python 3.11 is now the default, verify compatibility with your environment
- Qwen3 models in the LLM list have been replaced with Qwen3.5
New Features
- VibeVoice LoRA finetuning with comprehensive parameter UI
- Stop Training button to terminate subprocesses cleanly
- Qwen3.5 models added to the LLM list
- VibeVoice Streaming 0.5B with 7 built-in preset voices
- Optional DeepFilterNet for audio denoising
Improvements
- LoRA loading now auto-corrects PEFT task_type for better compatibility
- Trained model caching now includes smart reload on checkpoint change
- Audio denoising options via optional DeepFilterNet installation
Bug Fixes
- LoRA loading with PEFT task_type auto-correction
Upgrade Notes
Update your Python environment to 3.11 or later. Update your LLM configuration to use 'Qwen3.5' instead of 'Qwen3'. Note that DeepFilterNet is an optional dependency for audio denoising.
The Two Releases
Version 1.9.0 - CUDA Graphs Acceleration & Multi-GPU Support
Mar 1, 2026· 8 months ago
This release prioritizes performance optimization by integrating Faster-Qwen3-TTS for CUDA graph acceleration and introduces multi-GPU support to allow subsystems to run on separate hardware.
Full release notesVersion 1.10.0 - VibeVoice Lora Support, VibeVoice Preset Voices ( 0.5B model), Qwen3.5 LLM support
Mar 5, 2026· 8 months ago
A major update focusing on Voice AI training and LLM capabilities, specifically adding VibeVoice LoRA finetuning and Qwen3.5 support. It also includes significant UI refinements, Python 3.11 as the default, and audio denoising options.
Full release notes