ERNIE 5.0
Natively autoregressive omni-modal model trained from scratch on text, image, video and audio with a 2.4T-parameter ultra-sparse MoE (about 3% active).
Unifies understanding and generation with next-group-of-tokens prediction and elastic 'once-for-all' training that yields sub-models of varied depth and expert count. Tech report arXiv 2602.04705. Reached 1460 on LMArena text (#8) in January 2026.
- Date
- Thursday, 13 November 2025
- Lab
- Baidu
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | 2.4T <3% activated per token | company |
| LMArena text score (ERNIE-5.0-0110) | 1460, #8 globally 2026-01-15, Baidu report of LMArena | independent |
2025-11-13 is the Baidu World announcement (InfoWorld); preview checkpoints ran on LMArena from 2025-10. Baidu dropped the 'Preview' label by 2026-01-15; I did not find a dated official-release post. Proprietary, built on PaddlePaddle.
Sources
- www.infoworld.com/article/4089778/baidu-launches-new-generation-of-ernie-ai.html
- ernie.baidu.com/blog/posts/ernie5.0/
- arxiv.org/abs/2602.04705
- ernie.baidu.com/blog/posts/ernie-5.0-0110-release-on-lmarena/
This record was checked against its sources on 6 October 2026. How we check