Models & LabsEurope
Baidu releases ERNIE 5.5 Flash base model weights under a restricted license
Baidu on Nov. 5 released the weights for ERNIE-5.5-Flash-Base, a 280-billion-parameter mixture-of-experts language model that uses 14 billion parameters for each token. Its license bars use in the European Union, the United Kingdom and South Korea, and it requires Baidu's written approval once a user's related revenue passes $50 million a year.
The weights are on Hugging Face in FP8 and BF16 formats and are mirrored on ModelScope. The model is the raw checkpoint from the end of pre-training, before any instruction tuning. Baidu pitches it as a starting point for companies and researchers who want to adapt it to their own domains. The company says a post-trained version of ERNIE 5.5 Flash will follow.
The license, version 1.0 and dated Nov. 5, grants rights to run, modify, fine-tune and redistribute the model only inside its defined territory, which is everywhere except the EU, the UK and South Korea. Using the model or its outputs outside that territory is not covered by the grant.
Research, education, evaluation and personal use are free and need no notice. Commercial use is also free up to $50 million of annual revenue from products that rely on the model, measured over the trailing 12 months and added up across entities under common control. Commercial users in that band have to send Baidu a one-time notice before they start and display the words Powered by ERNIE in their product or documentation. They do not have to wait for Baidu to reply.
Above $50 million, users need Baidu's written authorization before any commercial use, and any fees are set in a separate agreement. A user that crosses the line without approval has to stop until approval takes effect. Applying or negotiating does not count as permission.
In practice, a developer or a smaller company can build on the model for free. A large company would need a deal with Baidu, and nobody in the three excluded markets can use it under this license.
Baidu's ERNIE 5.5 Flash Base opens weights with territory and revenue gates
Model scale, architecture and commercial license terms from the Nov. 5, 2026 release
| Item | Detail |
|---|---|
| Total parameters | 280 billion |
| Active per token | 14 billion |
| Expert layers | 42 layers; each picks 10 of 512 experts |
| Attention cache | One cache entry per 128 tokens |
| Training tokens | 33.5 trillion |
| Training context | Up to 128,000 tokens |
| Serving context | Supports 1 million tokens |
| Territory | Worldwide except EU, UK and South Korea |
| Commercial to $50m | Free with one-time notice; show Powered by ERNIE |
| Commercial above $50m | Needs Baidu written approval first |
The design aims to cut the cost of computing per token and of long prompts. Each of the model's 42 expert layers picks 10 of 512 experts for every token. Most of its attention layers keep one cache entry per 128 tokens rather than one per token, which shrinks the memory needed for long inputs. Baidu trained it from scratch on 33.5 trillion tokens with context up to 128,000 tokens. It says the model can be fine-tuned and served with context of about 1 million tokens.
Baidu supports serving through the open-source SGLang and vLLM engines, including an official SGLang container and its own vLLM plugin. The license does not commit Baidu to provide hosted inference, compute, support or future versions. Those require a separate agreement.
The company published benchmark results comparing the model with three base models of similar size from other labs, all run under the same test setup. It reported strong results on exams, knowledge, math and science, reasoning, multilingual, code and long-context tests. Baidu ran the tests itself.