Baidu releases ERNIE-5.5-Flash-Base weights under its own licence
A 280B-parameter base checkpoint with 14B active, pre-trained on 33.5T tokens; the licence is free up to US$50m in related revenue and excludes the EU, UK and South Korea.
- Models & capabilities
- Open weights & ecosystem
- Minor
Baidu published ERNIE-5.5-Flash-Base, a sparse mixture-of-experts model with 280 billion total parameters and 14 billion active per token, pre-trained from scratch on 33.5 trillion tokens. It is a base checkpoint rather than a finished chat model: the weights as they stood at the end of pre-training, with no mid-training or post-training. Baidu said developers can use it as a starting point for domain adaptation, instruction tuning or reinforcement-learning research. It is on Hugging Face in FP8 and BF16 formats, and Baidu said it supports a context of up to 1 million tokens for fine-tuning and inference, having been trained at up to 128,000.
Baidu compared it with the base versions of DeepSeek-V4-Flash, Tencent’s Hy3-preview and MiMo-V2.5, saying all four were run with the same harness, prompts and few-shot examples. The company’s own results show ERNIE ahead on most rows, for example MMLU 93.3 against 87.5 to 89.4 for the others, and MMLU-Pro 78.4 against 65.7 to 69.2. Competitors scored higher on some tasks, including ChineseSimpleQA and FullStackBench. The comparison is Baidu’s, and the release page cites no independent replication.
The licence is a custom “ERNIE 5.5 Flash Base Model License Agreement v1.0”. Non-commercial use is free. Commercial use is free of model licence charges for users whose related annual revenue is up to US$50 million, who must display “Powered by ERNIE” and send Baidu a one-time notice; above that, separate written authorisation and fees are required. The licence’s territory is everywhere except the EU, the UK and South Korea, and use or distribution of the model, derivatives and outputs outside that territory is barred. It also bars using outputs to improve other AI models, and its acceptable-use policy excludes military applications.
Four days later Baidu also described ERNIE Evolve, an automated-research system in which Director, Selector, Executor, Evaluator and Reviewer roles run experiments on a code repository. Baidu reported that it beat the best public results it cited on NanoChat AutoResearch and NanoGPT Speedrun tracks, and that its changes to unreleased ERNIE-5.5-Flash inference passed production validation. Those results are self-reported, though the post says outputs for the public benchmarks are on GitHub.