Baidu has released open weights for a new ERNIE model for the first time in 16 months, and the licence is the most interesting part for anyone reading this in London, Paris or Seoul. ERNIE-5.5-Flash-Base went up on Hugging Face and ModelScope on 5 November. Under the ERNIE 5.5 Flash Base Model License Agreement, the permitted “Territory” is the whole world except the European Union, the United Kingdom and South Korea.
What Baidu released
It is a sparse mixture-of-experts text model with 280B total parameters and 14B active per token. It was pre-trained from scratch on 33.5T tokens with a training context of up to 128K and supports 1M tokens for fine-tuning and inference. It is a base checkpoint, taken at the end of pre-training, with no instruction tuning or RL. It isn’t a chatbot, and Baidu says a post-trained ERNIE 5.5 Flash will follow. FP8 and BF16 weights are provided, with SGLang and vLLM recipes.
Baidu’s own benchmark table compares it with DeepSeek-V4-Flash-Base, Tencent’s Hy3-preview-Base and Xiaomi’s MiMo-V2.5-Base. ERNIE comes first on 24 of 30 rows, including MMLU-Pro at 78.4 against DeepSeek’s 69.2 and HumanEval+ at 82.3 against 69.5. DeepSeek still wins ChineseSimpleQA and the HELMET reranking test. As MarkTechPost and others noted, these are Baidu’s own results from Baidu’s own pipeline, with no independent verification yet.
For context, the last open ERNIE weights were the ERNIE 4.5 family in June 2025, released under Apache 2.0. Baidu didn’t publish weights for ERNIE 5.0 or 5.1. The full timeline is on our ERNIE Model Updates page, which we’ll update with this release.
The licence is not Apache 2.0
The new licence reads more like a commercial contract than an open-source licence. The terms that matter most:
- No UK, EU or South Korea. Section 5.3 says the weights, derivative models and their outputs must not be used, reproduced, modified, distributed or displayed outside the Territory. Using it outside the Territory is also the first item on the acceptable use policy. A UK agency can’t legally run this model, fine-tune it or show its outputs to clients under this licence.
- A US$50m revenue line. Commercial use is free, after a one-time notice to Baidu, if related products earn US$50m a year or less. Above that you need Baidu’s written authorisation first, and revenue is counted across the whole corporate group.
- “Powered by ERNIE” must be displayed on any commercial product or service, or in its API documentation if there is no user interface.
- No training other models. You can’t use the weights or outputs to improve any non-ERNIE model.
- Label machine-generated content. Item 12 of the acceptable use policy bans publishing generated material in public without clear, prominent identification that it is machine-generated, and it names articles and posts specifically.
Baidu hasn’t explained why these three markets are excluded. The EU AI Act’s obligations for general-purpose models, Korea’s AI Basic Act and UK copyright and online safety rules are all plausible reasons, but that is our reading, not Baidu’s.
What it means for search and SEO teams
For most readers outside China, the practical answer is simple: if you are in the UK, EU or Korea, you can’t use this model, so don’t build a content or classification pipeline on it. Elsewhere, any team using ERNIE output for published pages takes on a contractual duty to label it as machine-generated. That is stricter than anything Google asks for.
The bigger story is Baidu’s approach. A week after it cut its Search AI API to ¥4.9 per 1,000 calls, it is also competing with DeepSeek and Qwen on open weights, but on terms that keep control of large commercial users and certain regions. Baidu World is on 12 November, and a post-trained ERNIE 5.5 Flash, plus any change to the territory clause, is what we’ll be watching for.