xtuner
XTuner is an open-source training framework optimized for fine-tuning and pre-training ultra-large mixture-of-experts (MoE) models, with support for multimodal learning and reinforcement learning. It claims efficiency improvements over traditional 3D parallel training for models ranging from 200B to 1T parameters.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Repository | InternLM/xtuner |
| Owner | InternLM |
| Primary language | Python |
| License | Apache-2.0 — OSI-approved |
| Stars | 5.2k |
| Forks | 429 |
| Open issues | 342 |
| Latest release | v0.2.0 (2025-07-11) |
| Last updated | 2026-07-03 |
| Source | https://github.com/InternLM/xtuner |
What xtuner is
XTuner V1 is a PyTorch-based distributed training engine featuring dropless expert parallelism, long-sequence support via memory optimization and DeepSpeed Ulysses integration, and FSDP-based throughput claims exceeding traditional 3D parallelism for 200B+ MoE models. Core algorithms include multimodal pre-training, supervised fine-tuning, and GRPO reinforcement learning.
Get the xtuner source
Clone the repository and explore it locally.
git clone https://github.com/InternLM/xtuner.gitcd xtuner# follow the project's README for install & configurationNeed it deployed, integrated, or customized instead? DEV.co ships production installs.
Best use cases
Implementation considerations
- Requires proficiency in PyTorch distributed training, DeepSpeed/FSDP configuration, and multi-node cluster orchestration; expect significant DevOps investment for production setups.
- Model support matrix shows staged readiness: Qwen3-Dense and Intern models fully verified on GPU/NPU; Deepseek V3, Kimi K2, and GPT-OSS remain in-progress on Ascend.
- Long-sequence training (64k tokens) requires careful memory budgeting and validation; dropless expert parallelism reduces but does not eliminate load-balancing concerns at scale.
- Roadmap includes algorithm additions (MPO, DAPO, agentic RL); current stable features are multimodal pre-training, supervised fine-tuning, and GRPO; beta/unreleased features should be validated before production use.
- Integration with inference engines (LMDeploy confirmed, vLLM/SGLang pending) is necessary for deployment; training-to-inference handoff workflows are not yet fully documented.
When to avoid it — and what to weigh
- Small or Dense-Only Model Training — XTuner V1 is engineered specifically for ultra-large MoE architectures; training smaller dense models may not justify the complexity or resource overhead.
- Immediate Production Deployment Required — Latest release (v0.2.0, July 2025) is recent; only select model architectures (Qwen3-Dense, Intern S1/VL) are marked fully ready on GPU/NPU. Deepseek V3 and Kimi K2 support remain work-in-progress.
- Minimal DevOps and Infrastructure Expertise — Requires proficiency in distributed PyTorch, DeepSpeed configuration, multi-node orchestration, and Ascend/NVIDIA GPU provisioning; not beginner-friendly for training infrastructure setup.
- Vendor Lock-in Sensitivity — Strong alignment with InternLM ecosystem (InternS1, InternVL) and Ascend NPU optimization; less mature support for other proprietary or emerging accelerator architectures.
License & commercial use
Apache License 2.0 (Apache-2.0). This is a permissive, OSI-approved open-source license permitting commercial use, modification, and distribution under stated conditions (attribution, notice of license). README explicitly states the same.
Apache 2.0 permits commercial use. However, you must: (1) retain attribution and license notices, (2) disclose modifications, (3) include a copy of the license. Additionally, README states 'Please also adhere to the Licenses of models and datasets being used'—dependent models (Qwen3, Deepseek V3, Kimi K2, InternS1, InternVL) may have different license terms; verify each separately before commercial deployment.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Limited |
| License clarity | Clear |
| Deployment complexity | High |
| DEV.co fit | Good |
| Assessment confidence | High |
No explicit security audit, hardening practices, or vulnerability disclosure process mentioned in provided data. XTuner integrates DeepSpeed and PyTorch, inheriting their security posture. Training on sensitive datasets requires network isolation and access controls at the cluster level—ensure data residency and privacy compliance align with jurisdictional requirements. No built-in model audit or contamination detection for training data.
Alternatives to consider
Megatron-LM (NVIDIA)
Mature 3D parallel training framework for large language models; more stable but reportedly lower throughput than XTuner V1 for 200B+ MoE models per README benchmarks.
DeepSpeed + Custom FSDP (Microsoft)
Lower-level, modular approach; gives fine-grained control but requires custom engineering for MoE-specific optimizations and long-sequence handling.
Torchtitan (PyTorch)
PyTorch-native distributed training platform; simpler API but less specialized for ultra-large MoE architectures; XTuner V1 roadmap acknowledges Torchtitan as inspirational foundation.
Build on xtuner with DEV.co software developers
XTuner V1 offers production-grade distributed training for ultra-large models. Evaluate your cluster architecture, verify hardware compatibility (GPU/Ascend), and review deployment guides with your infrastructure team.
Talk to DEV.coRelated on DEV.co
Explore the category and the services that help you build with it.
xtuner FAQ
Can I use XTuner V1 for dense model training (non-MoE)?
Is Ascend NPU support production-ready?
What is the minimum cluster size to run XTuner V1?
How does XTuner V1 handle expert load imbalance during training?
Work with a software development agency
Need help beyond evaluating xtuner? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and ai frameworks integrations — and maintain them long-term.
Ready to Scale Your MoE Model Training?
XTuner V1 offers production-grade distributed training for ultra-large models. Evaluate your cluster architecture, verify hardware compatibility (GPU/Ascend), and review deployment guides with your infrastructure team.