LlamaFactory releases v0.9.5: supports Qwen3.5/3.6, Gemma 4 and Transformers v5
LlamaFactory v0.9.5 adds major support for Qwen3.5/Qwen3.6/Gemma4 models and is compatible with Transformers v5.
LlamaFactory v0.9.5: Catching up with the latest batch of models and moving towards Transformers v5
LlamaFactory releases v0.9.5, the official title directly addresses the topic - "Qwen3.5/3.6, Gemma 4, Transformers v5". This version does what the fine-tuning framework should do most: connect the latest open source models as soon as possible and keep up with the major versions of the upstream training stack.
Newly accessed model
| Categories | Models |
|---|---|
| Mainly recommended new | Qwen3.5, Qwen3.6, Gemma4 |
| More support | LiquidAI LFM2.5 / LFM2.5-VL, Phi-4-mini, Youtu-VL, MiniCPM-o-4.5, GLM-4.7-Flash, GLM-OCR |
The v1 training stack is also continuing to be improved, covering FSDP2, DeepSpeed, quantization, batch generation, etc.
Framework Compatibility: Keep up with Transformers v5
v0.9.5 is adapted to Transformers v5, and dependencies such as peft and deepspeed are updated accordingly. For teams who want to make fine-tuning on Qwen3.5/Gemma4, this version is basically ready for use out of the box; it also ensures synchronization with the upstream ecosystem.
Don’t forget several structural changes starting from v0.9.4
Before upgrading, please pay attention to the changes carried over from v0.9.4: the warehouse has been renamed LlamaFactory; requires Python 3.11+; the installation is migrated to uv (uv pip install llamafactory).
Check before upgrading
- First confirm that Python ≥ 3.11, and install it using uv according to the official instructions.
- Verify the compatibility of existing training scripts under Transformers v5 in an isolated environment before switching.
- The itemized changes and list of supported models are subject to the official release notes.
Reviews