LLaMA Factory

LLaMA Factory is an open-source fine-tuning framework with 75,000+ stars on GitHub that unifies training for more than 100 LLMs and vision-language models behind one consistent interface.

Overview

Fine-tuning used to mean a bespoke script for every model family. LLaMA Factory — published alongside an ACL 2024 paper — replaced that with one framework: pick a model (Llama, Qwen, DeepSeek, Gemma, ChatGLM and 100+ others), pick a method (full fine-tuning, LoRA, QLoRA, RLHF, DPO and more), point it at your data, and train. It supports instruction tuning, agent tuning and multimodal models, with memory-saving techniques like quantization and gradient checkpointing that let serious training happen on modest hardware. A web UI (LLaMABoard) makes configuration visual for those who prefer clicks to YAML. The project is Apache-2.0 licensed, maintained by an active research community, and updated continuously — repository activity in October 2026 — and it remains one of the most-used on-ramps to custom model training.

Key Features

• 100+ models – LLMs and VLMs from every major open family in one framework.
• Every major method – SFT, LoRA, QLoRA, RLHF, DPO and agent tuning.
• Memory efficient – quantization, FlashAttention and checkpointing for consumer-grade GPUs.
• Web UI included – LLaMABoard for visual configuration, monitoring and evaluation.
• Data flexibility – standard dataset formats plus templates for dozens of model families.
• Evaluation and export – benchmark trained models and merge adapters for deployment.

💳 Pricing

LLaMA Factory is completely free and open source under the Apache License 2.0. Your only cost is your own compute, local or rented.

Who Is It For

Researchers, engineers and advanced hobbyists who fine-tune open models — for domain adaptation, instruction following, agent behavior or research — and want one proven framework instead of per-model scripts.