LLaMA-Factory

Fine-tune 100+ models with a zero-code command line or web UI, covering LoRA, QLoRA, DPO and multimodal training.

  • Fine-tuning
  • Apache 2.0 · Open source
  • Windows, Linux
  • Verified

LLaMA-Factory is an open-source framework for fine-tuning more than 100 large language models and vision-language models, and its paper appeared at ACL 2024. You work through a zero-code command line or through LlamaBoard, a web UI built on Gradio. The README lists supported models such as LLaMA, Mistral, Qwen3, DeepSeek, Gemma and GLM, and supported training approaches.

According to the README, methods include continued pre-training, supervised fine-tuning including multimodal data, reward modeling, PPO, DPO, KTO and ORPO. For limited hardware it offers 16-bit full tuning, freeze-tuning, LoRA and 2 to 8-bit QLoRA through AQLM, AWQ, GPTQ, LLM.int8, HQQ and EETQ, and it can use FlashAttention-2, Unsloth, Liger Kernel and KTransformers. Experiment monitors include TensorBoard, Wandb and MLflow, and inference is available through an OpenAI-style API, a Gradio UI and a command line with a vLLM or SGLang worker.

The README has separate notes for Windows users, who must install the CUDA build of PyTorch themselves, and for Ascend NPUs and AMD GPUs. The license is Apache 2.0, and the latest release is v0.9.5 from May 2026. A fine-tuned model still carries the license of its base model.

Good for

  • Fine-tuning without writing training code
  • Trying LoRA and QLoRA on one GPU through a web UI
  • One tool for many model families
  • Learning how fine-tuning works with a visual front end

Watch out for

  • Windows users must install the CUDA build of PyTorch themselves, and QLoRA needs bitsandbytes
  • Check the supported-models list in the README for your model before you start
  • A fine-tuned model inherits the license of its base model, including any revenue limits

Links