LlamaFactory v0.9.5 is out. The headline additions are primary support for Qwen3.5, Qwen3.6, and Gemma 4 models, plus compatibility with Transformers v5. If you are building fine-tuning pipelines on any of these models, this is the release to move to.
The FP8 story gets a concrete fix here. The release adds Transformer Engine backend support, resolving an earlier FP8 issue. If you have been running into FP8 training instability, this patch is worth pulling immediately.
Several new model families land in this release. Microsoft's Phi-4-mini is now supported. LiquidAI's LFM2.5 and its vision-language variant LFM2.5-VL are both included (the LFM template was also renamed to LFM2 for consistency). Youtu-LLM-2B and HY-MT round out the new model additions.
On the loss side, EAFT loss support has been added. If your fine-tuning work involves specialized alignment objectives, this is a new option available without external patching.
Two robustness fixes are worth noting for teams with varied deployment environments. The release now handles an empty architectures field in config.json without breaking, and it adds a PyTorch version warning for Conv3D, which matters if you are working with multimodal or video models on older PyTorch builds.
The v1 API layer is also progressing. This release adds an init plugin, a CLI sampler, a batch generator, and renderer unit tests. These are internal scaffolding changes, but they signal that a more structured inference API is taking shape inside the framework.
CI got attention too. CUDA CI caching was improved, which speeds up the contributor workflow and should make future releases faster to validate.
What to do today: If you are fine-tuning Qwen3.5, Qwen3.6, or Gemma 4 models, upgrade to v0.9.5 now. Verify your config.json files do not rely on a populated architectures field, and test your FP8 setup with the Transformer Engine backend enabled. If you are running Phi-4-mini or any LiquidAI LFM2.5 model, this is the first release with native template support, so it is the right baseline to build on.