Baidu open-sources a 280B model, but not for the EU, UK or Korea

Baidu’s ERNIE team released ERNIE 5.5 Flash Base on 5 November, open weights for a 280-billion-parameter mixture-of-experts model that activates 14 billion parameters per token. It was pre-trained from scratch on 33.5 trillion tokens with a 128K training context, and supports a 1-million-token context for fine-tuning and inference. Weights are on Hugging Face in FP8 and BF16, with deployment support for vLLM and SGLang. Baidu reports 93.3% on MMLU and 82.3% on HumanEval+. This is a base checkpoint only, ending before mid-training or post-training, meant as a starting point for fine-tuning and reinforcement-learning work rather than a drop-in chat model.

The licence is the new catch. Unlike ERNIE 4.5’s Apache 2.0 terms, the ERNIE 5.5 Flash Base Model License Agreement v1.0 defines its “Territory” as everywhere except the European Union, the United Kingdom and South Korea, and bars using, reproducing or distributing the model outside it. Commercial use is otherwise free up to $50 million in annual revenue from products built on it, with just a notice email to Baidu. Past that threshold, you need separate written authorisation first.

Why it matters: this is a frontier-scale base model you could fine-tune and self-host, but the licence itself decides who gets to build on it. Anyone in the EU, UK or South Korea is excluded outright, and anyone scaling past $50 million in related revenue needs Baidu’s sign-off before they can keep using it commercially.