LlamaFactory
UnknownLlamaFactory is a toolkit for fine-tuning 100+ open LLM and vision-language architectures, including LLaMA, Qwen, DeepSeek, Gemma, and GLM. It covers continued pre-training, supervised fine-tuning, reward modeling, and preference or RL methods such as PPO, DPO, KTO, and ORPO, across full, LoRA, and QLoRA tuning. A CLI and the Gradio-based LLaMA Board web UI run the same pipeline without custom code, and the project appeared as an ACL 2024 paper.
The project renamed itself from LLaMA-Factory to LlamaFactory in 2026, and the old repository URL redirects to hiyouga/LlamaFactory.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public(hiyouga/LlamaFactory)
- core features withheld
- no
Apache-2.0, an OSI license, with the full source public including the web UI, and no features held back for a paid tier.
- https://github.com/hiyouga/LLaMA-Factory recorded 2026-09-09
Redirects (301) to the renamed repository https://github.com/hiyouga/LlamaFactory (same owner, repo id 646410686, created 2023-05-28). The rendered README states "This repository is licensed under the Apache-2.0 License." The top-level tree is .ai, .claude, .github, assets, data, docker, docs, examples, requirements, scripts, src, tests, tests_v1 plus root files including LICENSE — no enterprise, commercial, or paid-tier directory, and src/ ships the whole toolkit including the CLI and the Gradio-based LLaMA Board web UI. Not archived; sidebar reports 73,920 stars and 9,042 forks.
- https://pypi.org/project/llamafactory/ recorded 2026-09-23
Project page for "llamafactory 0.9.5" lists License expression Apache-2.0 and points Homepage/Repository at https://github.com/hiyouga/LLaMA-Factory, confirming the published package matches the GitHub source.
- https://raw.githubusercontent.com/hiyouga/LlamaFactory/main/LICENSE recorded 2026-09-23
The full standard Apache License, Version 2.0 text — definitions through the appendix boilerplate — with no additional restriction, exception, or commons-clause layered on top.
Adoption
2 medium confidenceLlamaFactory's PyPI downloads climbed to a sharp peak in mid-2026 and have fallen every month since, so its score reflects current package-download volume rather than the peak. The package is a floor rather than a full measure, since many people run the project from a git clone and its bundled web UI rather than a PyPI install.
- https://api.github.com/repos/hiyouga/LLaMA-Factory/releases recorded 2026-08-09
30 releases carrying 5,242 asset downloads in total, so no meaningful distribution channel outside PyPI
- https://github.com/hiyouga/LLaMA-Factory recorded 2026-08-09
"73939 users starred"
- https://pypistats.org/api/packages/llamafactory/overall?mirrors=false recorded 2026-09-24
daily download series; summed by month it gives 30,043 (Apr), 233,960 (May), 779,461 (Jun), 100,640 (Jul), 23,511 (Aug) and 11,990 (Sep, through the 23rd) for 2026
- https://pypistats.org/api/packages/llamafactory/recent recorded 2026-09-24
"last_month":19183
Capability
4 medium confidenceLlamaFactory fine-tunes more than 100 LLM and vision-language models with full-parameter tuning, LoRA and QLoRA, preference and reinforcement methods such as PPO and DPO, and a CLI or the zero-code LLaMA Board web UI. It trails Megatron-LM on scale: it reaches large models through FSDP with QLoRA and an optional Megatron-core backend, while Megatron-LM composes tensor, pipeline, context and expert parallelism and reports training a 462B-parameter model on thousands of GPUs.
- https://raw.githubusercontent.com/hiyouga/LlamaFactory/main/README.md recorded 2026-09-28
The README headline reads "Easily fine-tune 100+ large language models with zero-code" through a CLI and the LLaMA Board web UI, and lists "(Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO" plus "16-bit full-tuning, freeze-tuning, LoRA". Its news notes "fine-tunes a 70B model on 2x24GB GPUs" with FSDP+QLoRA and "We support Megatron-core training backend with" an external adapter.
Verified 2026-09-09