XTuner

Shanghai AI Laboratory's InternLM team maintains XTuner, a toolkit that began as an efficient fine-tuning library offering QLoRA, LoRA, full-parameter, and DPO training across LLMs and vision-language models, and has since evolved into a next-generation engine targeting ultra-large mixture-of-experts models. The current V1 line focuses on multimodal pretraining and supervised fine-tuning plus GRPO reinforcement learning for models including Intern S1, InternVL, Qwen3 dense and MoE variants, GPT-OSS, DeepSeek V3, and Kimi K2, with MPO, DAPO, and multi-turn agentic RL on the roadmap. It advertises dropless training and 64k sequence lengths at the 200B MoE scale. Installation is pip install xtuner, a Dockerfile is provided, NVIDIA GPUs or Ascend NPUs are assumed, and documentation lives at xtuner.readthedocs.io. The Apache-2.0 license keeps commercial use open, and around 5,200 GitHub stars track a project that serves as the official training path for the InternLM ecosystem while pushing into large-scale MoE territory.