LlamaFactory v0.9.4 lands with a set of breaking changes that affect every existing installation. The project has also renamed itself from LLaMA-Factory to LlamaFactory, so update any internal references accordingly.
Python 3.9 and 3.10 are no longer supported. The framework now requires Python 3.11 to 3.13. If your training environment is pinned to an older runtime, you need to upgrade before pulling this release.
The package manager has changed. LlamaFactory has migrated from pip to uv. The new install command is uv pip install llamafactory. Drop the old pip invocation from your CI scripts and Dockerfiles.
On the feature side, the headline addition is Orthogonal Fine-Tuning (OFT). The technique is described in this paper and gives practitioners another fine-tuning method alongside existing LoRA-style approaches. A complementary addition is Semantic Initialization for newly added tokens, which should matter any time you extend a model vocabulary for a domain-specific task.
For teams running large-scale training, Megatron-LM support arrives via MCoreAdapter. This opens up distributed training configurations that were not previously accessible through LlamaFactory. On the inference side, the new KTransformers backend gives you an additional runtime option at serving time.
Other notable additions include FP8 training, DeepSpeed AutoTP, efficient NPU fused kernels, support for Transformers v5, TRL 0.24, and the MPO alignment algorithm.
The model roster grew substantially. New integrations include Falcon H1, Kimi-VL, GLM-4.5V, Gemma3n, Granite4, Qwen3-2507, MiniCPM-V 4.0 and 4.5, InternVL-3.5, ERNIE-4.5-Text, ERNIE-4.5-VL, Qwen3-VL, Qwen3-Omni, Hunyuan-mt, GLM-4.6V, Ministral 3, VibeThinker, MiMo-V2-Flash, MiniMax-M1, and MiniMax-M2. If you work with any of these, the integration is community-contributed and ready to test.
The project also launched an official blog at blog.llamafactory.net, which is worth bookmarking for roadmap signals.
What to do today: Audit your Python version and package manager setup before upgrading. Bump to Python 3.11 or higher, swap pip for uv, update your install command to uv pip install llamafactory, and run your fine-tuning pipeline on the new version. If you have been waiting for Megatron-LM support or a KTransformers inference path, this is the release to test those configurations.