LLaMA-Factory

LLaMA-Factory is a unified framework for fine-tuning more than a hundred open language and vision models, with a simple CLI and web UI. Machine-learning engineers use it to customize open models on their own data.

LLaMA-Factory is an open-source framework for fine-tuning large language and vision-language models without writing training code. It supports 100+ open models (Llama, Qwen, Gemma, Mistral, DeepSeek, GLM, Phi, and many multimodal models) and a wide range of training approaches: continued pretraining, supervised fine-tuning, reward modeling, and preference methods like PPO, DPO, KTO, and ORPO. It covers full-parameter, freeze, LoRA, and 2-to-8-bit QLoRA tuning, plus efficiency tricks such as FlashAttention-2, Unsloth, and Liger Kernel. You can drive it entirely from a config-based CLI or from LLaMA Board, a Gradio web UI, which is where the "zero-code" label comes from. It integrates experiment trackers (TensorBoard, Weights & Biases, MLflow, SwanLab) and faster inference via vLLM or SGLang with an OpenAI-style API, and it can export checkpoints, including an Ollama modelfile. Released under Apache-2.0, it is widely used by researchers and practitioners customizing open models, and it requires an NVIDIA (or compatible) GPU.