Baidu published the weights of ERNIE 5.5 Flash Base on Hugging Face, with a licence dated November 5, 2026.
The model is a sparse mixture of experts with 280 billion total parameters and 14 billion active per token. It was pre-trained from scratch on 33.5 trillion tokens. Training context was up to 128K tokens. The released configuration supports a 1 million token context for fine-tuning and inference.
It is a base checkpoint, taken at the end of pre-training and before any instruction or reinforcement training. Baidu positions it as a starting point for domain adaptation and research. FP8 and BF16 weights are available, with serving support in SGLang and vLLM.
Baidu reports MMLU-Pro of 78.4 against 69.2 for DeepSeek-V4-Flash-Base in its own evaluation pipeline. These are Baidu's figures.
The licence is the main news for European users. It grants rights only within a Territory defined as all areas worldwide other than the European Union, the United Kingdom and South Korea.
Inside the Territory, non-commercial use is free. Commercial use is free up to US$50 million in annual revenue from related products, after a one-time notice to Baidu and with "Powered by ERNIE" attribution. Above US$50 million, prior written authorization is required. Revenue of affiliated companies is counted together.
Why this matters
Open weights do not mean open everywhere. A company in Bulgaria, Germany or Austria has no licence to use this model, even for internal tests. Every open-weight model needs a licence check for territory, revenue caps and attribution before anyone downloads it.
Sources: ERNIE 5.5 Flash Base model card, licence agreement.
This article is not legal advice.
