Model components / Base / pretrained models

ERNIE 5.1

Baidu's flagship fifth-generation LLM (released May 2026): a Mixture-of-Experts model with elastic depth/width/sparsity, focused on text generation, reasoning, agentic tool-use, and creative writing. Baidu emphasizes cost efficiency - total params ~1/3 and active params ~1/2 of ERNIE 5.0, trained at ~6% of comparable models' pre-training cost. 128K context.

Proprietary/API-only (Baidu Qianfan API, ernie.baidu.com, AI Studio). Note ERNIE 4.5 was open sourced (Apache-2.0, June 2025) but the 5.x flagship line is closed. LMArena frontier (~#28 per mirrors; Baidu claims 5.1-Preview #13 Text / #1 Chinese).

Openness

1 high confidence

1.0

weights: closed
data: closed
code: closed
license: proprietary(Baidu Qianfan API)

No weights distributed for the 5.x flagship; closed API model.

https://huggingface.co/baidu recorded 2026-06-18
Baidu HF org lists only ERNIE 4.5 + Image open, no 5.x weights

Adoption

4 low confidence

4.0

Baidu's ERNIE/Wenxiaoyan assistant app reported ~200M MAUs around the 5.0 launch (app-level, not 5.1-specific); not on HF.

https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/ recorded 2026-06-18
ERNIE 5.1 release; API access

Capability

5 medium confidence

5.0

Frontier proprietary model; Arena rank disputed between mirrors (#28) and Baidu's own claim (#13 Text).

https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/ recorded 2026-06-18
official MoE design, benchmarks, training-cost efficiency

Unchanged since 2026-07-30 (last edited, not re-checked)