LLaMA-Factory

LLaMA-Factory is a local model training platform for developers and researchers who want to adapt pretrained models to their own tasks. Its web UI lets you fine-tune models without writing code.

• Language and multimodal models: Train LLaMA, Qwen3, Qwen3-VL, DeepSeek, Gemma, Mistral and LLaVA for dialogue, tool use, image understanding, video recognition or audio understanding.
• Training methods: Continue pre-training, fine-tune with labeled examples, or adjust responses through reward modeling and preference training methods such as DPO and PPO.
• Memory-saving fine-tuning: Choose LoRA or quantized QLoRA to reduce memory requirements, alongside full-weight training and training with selected weights frozen.
• Experiment tracking: Compare training runs with LlamaBoard, TensorBoard, Wandb or MLflow.
• Model inference: Use trained models through a Gradio interface or an OpenAI-style API, with support for Transformers, vLLM and SGLang inference backends.

The project is open source under Apache 2.0 and supports local fine-tuning, CUDA, and distributed training across multiple GPUs or machines. Hardware requirements depend on the model and training method. Estimated memory needs for a 7B model are about 6 GB with 4-bit QLoRA and 16 GB with LoRA. Optional acceleration integrations include FlashAttention-2, Unsloth and Liger Kernel.