Baidu releases ERNIE 5.5 Flash Base weights, under a licence that excludes the UK and EU

Baidu has published the weights of ERNIE-5.5-Flash-Base, which its release page, dated 5 November 2026, describes as "a sparse Mixture-of-Experts (MoE) language model with 280B total parameters, 14B of which are activated per token". The weights are on Hugging Face in FP8 and BF16 versions under a custom licence that, on its own terms, grants no rights in the UK, the European Union or South Korea.

What Baidu released

According to Baidu's page, the model was "pre-trained from scratch on 33.5T tokens" with a training context of up to 128K tokens, and supports a 1M-token context for fine-tuning and inference.

The checkpoint is a base model. Baidu says the weights are "the checkpoint at the end of pre-training, before any mid- or post-training", and pitches them as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. In practice, per Baidu's own documentation:

• SGLang and vLLM serve it "through the completions API".
• "The model expects a BOS token at the start of every input".
• Baidu writes: "This base model is only a starting point. The post-trained ERNIE 5.5 Flash will follow". It gives no date.

When the weights appeared

Baidu's page and the licence both carry 5 November 2026. Hugging Face's public API shows a slightly longer timeline: the FP8 repository baidu/ERNIE-5.5-Flash-Base was created on 21 October, its weight files were uploaded on 2 November and some re-uploaded on 3 November, the licence file was uploaded on 2 and 3 November, and the model card was edited several times on 5 November. The BF16 repository, also created on 21 October, received its weights in a single upload on 30 October. Both repositories are now public and not gated. The API does not record when a repository was switched from private to public, so we cannot say whether the files were downloadable before 5 November; the earliest Internet Archive captures of both model pages are from 5 November.

The licence

Hugging Face lists the licence as "other" (ernie-5.5-flash-base-model-license-v1.0), not a standard open-source licence. We read the ERNIE 5.5 Flash Base Model License Agreement v1.0 in full. Its main terms:

• Territory. Rights are granted "solely within the Territory", defined as "all geographical areas worldwide other than the European Union, the United Kingdom and South Korea". The licence adds that the materials, derivative models and their outputs "must not be used, reproduced, modified, distributed or displayed outside the Territory".
• Non-commercial use (research, education, evaluation, personal experimentation) is free, with no notice required.
• Commercial use is free where revenue from related products is at or below US$50 million a year, provided the user sends Baidu a one-time notice and displays "Powered by ERNIE". Above US$50 million, separate written authorisation is required first. Revenue is aggregated across companies under common control.
• Improving other models. Users may not use the model or its outputs "to improve any other artificial intelligence model or machine-learning system other than the Model or its Derivative Models". The licence counts a model distilled from ERNIE as a Derivative Model, so distillation is permitted but the result stays under this licence.
• Acceptable use. A 20-item policy bars, among other things, military use, high-stakes automated decisions in areas such as credit, employment and healthcare, and publishing generated material publicly without clear labelling as machine-generated.
• Disputes go to the courts of Hong Kong, under Hong Kong law.

Given these conditions, we describe the release as open weights rather than open source. On our reading, the licence would not meet the Open Source Initiative's definition, which says a licence "must not restrict anyone from making use of the program in a specific field of endeavor"; this one bars military use, for example.

Baidu's benchmarks

All figures are Baidu's own. It compares the model with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, which it calls "the latest open-source base models of similar size publicly available at the time of release", using the same harness and prompts for all four.

Baidu's table has 30 benchmark rows. ERNIE has the top score in 24, including MMLU-Pro (78.4 against 69.2 for DeepSeek), LiveCodeBench (40.1 against 33.5 for Hy3) and HumanEval+ (82.3 against 78.0 for Hy3). A rival leads in six:

• ChineseSimpleQA: DeepSeek 75.3, ERNIE 73.0
• BBH with chain-of-thought: Hy3 90.0, ERNIE 89.2
• INCLUDE-lite: DeepSeek 79.8, ERNIE 79.5
• FullStackBench: Hy3 56.7, ERNIE 54.6
• HELMET-Rerank: DeepSeek 60.0, ERNIE 56.2
• InfiniteBench: DeepSeek 48.0, ERNIE 46.5

Several margins in both directions are under one point. Baidu's page cites no independent evaluation.

Why it matters

On our reading, Baidu's figures put a 14B-active model ahead of similar-sized open base models on most of its own chosen tests, and releasing a pre-training-only checkpoint gives researchers a clean baseline, which is how Baidu frames it. On our reading, the licence matters as much as the scores for readers here: as written, its grant does not extend to the UK or the EU, so organisations in those places would need to settle that with Baidu before using the model.