#4545th of 60 in Training and fine-tuning
xtuner
Training engine for fine-tuning and RL of large MoE and dense LLMs
- Stars
- 5.2k
- License
- Apache-2.0
- Last commit
- Nov 2026
- Last release
- Jul 2025
- Language
- Python
5,079 on 21 Apr 20265,221 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
XTuner V1 is a PyTorch training engine for pre-training, instruction fine-tuning and reinforcement learning of large MoE models. It uses FSDP with optional DeepSpeed Ulysses sequence parallelism instead of 3D parallelism, and the README claims 200B MoE training without expert parallelism and support up to 1T parameters. It covers Intern S1, Intern VL, Qwen3, GPT OSS, DeepSeek V3 and Kimi K2 on NVIDIA GPUs and Ascend NPUs.
Who it is for: Teams training or fine-tuning large MoE LLMs on GPU or Ascend clusters
Strengths
- FSDP-based MoE training without expert parallelism at 200B scale, per the README
- Supports FP8 and BF16 on GPU and BF16 on Ascend NPU
- Includes multimodal pre-training, multimodal SFT and GRPO
- DeepSpeed Ulysses sequence parallelism for long sequences, with 64k claimed on 200B MoE
Weaknesses
- Last tagged release is July 2025, older than the V1 announcement
- MPO, DAPO and multi-turn agentic RL are still listed as coming soon
- GPT OSS, DeepSeek V3 and Kimi K2 are not yet supported on NPU
- Inference integration covers only LMDeploy; vLLM and SGLang are unchecked
What it needs
- GPU required
- Docker
- Needs PyTorch
- Models: Intern S1, Intern VL, Qwen3 Dense, Qwen3 MoE, GPT OSS
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Also in Training and fine-tuning
See all 60| Rank | Project | Score |
|---|---|---|
| 1 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 89 out of 100 |
| 2 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 3 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 87 out of 100 |
| #4 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #5 | megatron-lmDistributed training library for large transformer models on NVIDIA GPUs | 85 out of 100 |