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 have been waiting to fine-tune any of those models inside LlamaFactory, the blocker is now gone.
Beyond the flagship model additions, the release packs a lot of practical fixes and expansions.
FP8 training gets a meaningful upgrade. The Transformer Engine backend is now supported, correcting a gap in the previous FP8 implementation. If you are running FP8 workloads on compatible hardware, update and re-test your setup.
Several new model families land in this release. LiquidAI's LFM2.5 and LFM2.5-VL (the vision-language variant) are now supported. Microsoft's Phi-4-mini is in. Youtu-LLM-2B and the HY-MT model are included as well. Each of these was contributed by community members, which signals that the model coverage surface is expanding fast.
On the training side, support for EAFT loss has been added. This gives builders another loss function option when experimenting with fine-tuning strategies.
A few reliability and compatibility fixes are worth noting. The release handles an empty architectures field in config.json gracefully, which previously could cause silent failures when loading certain model configs. A PyTorch version warning for Conv3D was added to catch environment mismatches earlier. The ktransformers example config paths and templates were also corrected.
The v1 API layer is taking shape internally. This release adds an init plugin, a CLI sampler, a batch generator, and renderer unit tests. These are infrastructure pieces that point toward a more modular inference and serving layer, though they are not yet the primary interface for most users.
CI improvements land too. CUDA cache handling in CI was improved, which matters if you are running your own fork with GPU-based continuous integration.
What should you do today? If you are targeting Qwen3.5, Qwen3.6, or Gemma 4 for fine-tuning, upgrade to v0.9.5 and validate your training config. If you are running FP8 training, add the Transformer Engine backend to your setup and verify your existing workflows still behave as expected. And if you are loading community or custom model configs that omit the architectures field, you can now do so without workarounds.