LlamaFactory v0.9.5 ships primary support for Qwen3.5, Qwen3.6, and Gemma4 models, plus compatibility with Transformers v5. If you have been holding off on upgrading your Transformers installation, this release removes that blocker.
The headline additions are broad but practical. Qwen3.5 and Qwen3.6 cover the latest Qwen generation. Gemma4 support lands at the same time. Transformers v5 compatibility means the dependency chain can now move forward without breaking fine-tuning workflows built on LlamaFactory.
Beyond the flagship model families, the release adds support for several other models. Microsoft Phi-4-mini, LiquidAI's LFM2.5, LFM2.5-VL (a vision-language variant), Youtu-LLM-2B, and HY-MT all gain support in this version. That is a wide net cast in a single release, covering both text and vision-language use cases.
On the training side, EAFT loss support is now available. If your fine-tuning experiments require that loss function, you no longer need a workaround.
FP8 training also gets two fixes. Transformer Engine backend support is added for FP8, and a follow-up patch corrects remaining issues. For teams running FP8 on compatible hardware, the path is cleaner now.
A few smaller but operationally useful fixes round out the release. The framework now handles an empty architectures field in config.json without breaking, which matters when working with non-standard or community model checkpoints that omit that field. A PyTorch version warning is added for Conv3D, flagging compatibility issues before they become silent failures. The ktransformers example config paths and templates are corrected.
The v1 internals also move forward. This release adds an init plugin, a CLI sampler, a batch generator, and a renderer unit test. These are infrastructure pieces for the v1 API surface, not user-facing features yet, but they signal active development on that layer.
CI cache improvements for CUDA builds are included, which shortens iteration time for contributors running the test suite.
What to do today: If you are training or fine-tuning on Qwen3.5, Qwen3.6, or Gemma4 models, upgrade to v0.9.5 now. If you were pinned to an older Transformers version to avoid breakage, this release is your signal to unpin. Check your config.json files for any missing architectures fields if you have been seeing load errors on custom checkpoints. And if you run FP8 training with Transformer Engine, the backend support added here is worth testing in your pipeline.