XTuner
Shanghai AI LaboratoryXTuner is a training engine for large mixture-of-experts models. Version 1 targets dropless MoE training at up to 1T parameters, long-sequence training through memory optimization rather than sequence parallelism, and FSDP-based throughput for MoE models above the 200B scale. It is maintained by Shanghai AI Laboratory.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public(InternLM/xtuner)
- core features withheld
- no — no enterprise or ee path in the repository root and no paid tier described in the README
XTuner is released under Apache-2.0, an OSI license, with the full source public and no enterprise or paid tier described anywhere in the repository or README.
- https://raw.githubusercontent.com/InternLM/xtuner/main/LICENSE recorded 2026-10-06
LICENSE body, read in full: the stock Apache License 2.0 text, carrying no appended term or additional condition.
- https://raw.githubusercontent.com/InternLM/xtuner/main/README.md recorded 2026-10-06
README, read for the distribution and gating question: it describes installing and running the code from this repository and names no enterprise, paid or licence-gated build.
Adoption
1 medium confidenceXTuner's PyPI package has stayed at 0.2.0, uploaded 2025-07-11, while the repository released v1.0.1 on 2026-05-15, so a download count from the old package measures a version the project has left behind rather than current use. No stronger usage signal is available, since GitHub stars measure attention rather than use.
- https://api.github.com/repos/InternLM/xtuner recorded 2026-08-31
Repo metadata for InternLM/xtuner: stargazers_count 5187, archived false, pushed_at 2026-08-31 - the attention figure the note contrasts with the download count.
- https://api.github.com/repos/InternLM/xtuner/releases?per_page=3 recorded 2026-08-31
Newest three releases: v1.0.1 published 2026-05-15T06:57:47Z (prerelease true, draft false), v1.0.0rc0 of 2025-11-18, and v0.2.0 of 2025-07-11T05:43:37Z - the last release the registry matched, published the same day as the PyPI upload.
- https://pypi.org/pypi/xtuner/json recorded 2026-08-31
info.version 0.2.0, uploaded 2025-07-11T05:43:58Z, newest of 27 releases; nothing on a 1.x line has ever been published to the registry.
- https://pypistats.org/api/packages/xtuner/recent recorded 2026-08-31
last_month 1159, last_week 112, last_day 11 for the xtuner package. Recorded as a measurement of the trailing registry line, explicitly NOT as the basis of the reading.
Capability
4 medium confidenceXTuner's README claims frontier MoE scale -- 1T parameters, 200B without expert parallelism, 64k sequences without sequence parallelism -- but that throughput claim is stated in the README rather than a published scaling study. Megatron-LM offers composable parallelism across model families with published weak- and strong-scaling benchmarks that XTuner, specialized to MoE, does not match.
- https://raw.githubusercontent.com/InternLM/xtuner/main/README.md recorded 2026-10-07
Key Features section states dropless training of 200B MoE without expert parallelism and 600B with intra-node EP, 64k-sequence training on 200B MoE without sequence parallelism, DeepSpeed Ulysses support, MoE training up to 1T parameters, and FSDP throughput surpassing traditional 3D parallel schemes for MoE above 200B.
Verified 2026-10-06, except some scores, still awaiting re-confirmation