Back to Gap Map Model components / Base / pretrained models

ERNIE Flash

Baidu
open weights / Overall score: 3.1

ERNIE Flash is Baidu's line of smaller, efficient ERNIE models. Its open release, ERNIE 5.5 Flash Base, published on Hugging Face in October 2026, is a mixture-of-experts base model with 280B total and 14B active parameters, pretrained from scratch and released at the end of pretraining so developers can continue training it. The post-trained ERNIE 5.5 Flash Preview is offered only as a hosted model.

Scored on ERNIE 5.5 Flash Base, the open pretraining checkpoint. The API-only ERNIE 5.5 Flash Preview is left out of the openness reading, and Baidu's flagship line is the separate ernie entry.

Openness

3 high confidence
3.0
weights
open (FP8 and BF16 safetensors, ungated on Hugging Face)
data
closed (the pretraining data is described by category (web, code, long context), with no corpus, listing or mixture released)
code
partial (inference through SGLang and the vLLM-ERNIE plugin; no pretraining pipeline)
license
ERNIE-5.5-Flash-Base-Model-License-1.0 (commercial use free up to US$50 million in related annual revenue after a one-time notice, written authorization above it; the EU, the UK and South Korea are excluded)
api-only-tier
ERNIE 5.5 Flash Preview (the post-trained model, offered only as a hosted model with no Hub checkpoint, so excluded)

Baidu publishes the ERNIE 5.5 Flash Base weights on Hugging Face, ungated, under its own license, which allows commercial use without a fee up to US$50 million in related annual revenue and requires written permission above that. The license does not cover the EU, the UK or South Korea. Baidu describes its pretraining data only by category and publishes inference support rather than a training pipeline, which keeps it at open weights.

Adoption

1 high confidence
1.0

Hugging Face downloads summed across the FP8 and BF16 copies of ERNIE 5.5 Flash Base, released in late October 2026.

Capability

4 low confidence
4.0

Baidu's own comparison of base checkpoints, run under a single evaluation pipeline, puts ERNIE 5.5 Flash Base ahead of the base model behind Tencent's Hy3 on almost every row, and ahead of DeepSeek's and Xiaomi's similar-size bases on most. The results are vendor-run and cover base models only, with no independent leaderboard reading, so it is placed level with Hunyuan rather than above it.

  • https://huggingface.co/baidu/ERNIE-5.5-Flash-Base/raw/main/README.md recorded 2026-11-12

    Card evaluation table, "all evaluated with the same pipeline, prompts and few-shot settings": ERNIE-5.5-Flash-Base against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, with rows such as "MMLU-Pro (EM) | 5-shot | 78.4 | 69.2 | 66.9 | 65.7" and "GPQA (EM) | 5-shot | 60.2 | 55.9 | 50.2 | 51.4".

Verified 2026-11-12