XTuner is InternLM’s toolkit and next-generation training engine for fine-tuning large language models and multimodal models. It covers the full workflow from data preparation through supervised fine-tuning of models such as InternLM, Qwen, DeepSeek and Llama, scaling from a single GPU to clusters training ultra-large mixture-of-experts models. The Apache-2.0 project has more than 5,200 stars on GitHub and is under very active development in 2026.
XTuner began as an efficient fine-tuning toolkit โ supporting full-parameter training and parameter-efficient methods such as LoRA and QLoRA with DeepSpeed and FSDP backends, plus a large library of ready-made configs that made one-command fine-tuning practical on modest hardware. It has since grown into a broader training engine aimed at the newest model generations: the repository now highlights ultra-large MoE training, multimodal models such as InternVL and Qwen-VL variants, and reinforcement learning workflows alongside classic SFT. Data tooling, conversation-format handling and evaluation hooks are built in, so a team can go from raw text to a fine-tuned, chat-ready checkpoint without assembling separate scripts. Coming from the InternLM team at Shanghai AI Laboratory, XTuner is battle-tested on their own model releases and is one of the most widely used open toolkits in the Chinese LLM ecosystem, with extensive documentation and examples.
Free. XTuner is open source under the Apache-2.0 license.
ML engineers and research teams fine-tuning open-weight LLMs and multimodal models, especially those working with the InternLM ecosystem or training large MoE models.