XTuner is a Python-based training engine for ultra-large MoE models, detailing features and releases. It has 5172 stars and 349 open issues as of the latest data.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
XTuner is a Python-based training engine designed for ultra-large MoE models.
How it works
The repository describes an engine optimized for training MoE models at scale, with features such as dropless training, long sequence support, and efficiency improvements. It includes integration points with inference frameworks and data preparation tools. (No specific internal architecture details are provided in the provided text.)
Getting started
"## Getting started" is not included in the provided material. Therefore, this section is omitted.
Recent releases
- v1.0.1 bk main (2026-05-15)
- v1.0.0rc0 v1.0.0rc0 (2025-11-18)
- v0.2.0 XTuner Release V0.2.0 (2025-07-11): Added Support for Pre-trained RM and related bug fixes
- v0.2.0rc0 (2025-02-21): Supported FSDP2 and Contiguous Batching for RLHF, MiniCPM support
- v0.1.23 XTuner Release V0.1.23 (2024-07-22): Supported InternVL 1.5/2.0 finetune and related bug fixes
Traction
5172 stars, 437 forks, 349 open_issues (as raw counts; since inception: created 2023-07-11, last_push 2026-08-04)
Behind the repo
No linked startup/company information is provided in the material, so this section is omitted.
Caveats
- License: Apache-2.0
- Created: 2023-07-11
- Last push: 2026-08-04
- Language: Python
- Open issues: 349
- License and project details are provided; no additional age-related caveats beyond creation date are given in the material.