Baidu Releases Open Weights for ERNIE 5.5 Flash Base, a 280 Billion-Parameter Model

The release puts Baidu back into the contest among Chinese labs to publish large open models. That contest has made models from DeepSeek, Xiaomi and Tencent popular with developers and companies that want to run and adapt AI systems on their own infrastructure. By handing out a checkpoint from before instruction tuning, Baidu is targeting researchers and companies that want to do their own domain adaptation, instruction tuning and reinforcement learning, which is how the model card describes its intended use.

The model uses a mixture-of-experts design in which 14 billion of its 280 billion parameters are active for each token, keeping the computing cost of running it well below that of a dense model of the same size. According to the model card, Baidu pre-trained it from scratch on 33.5 trillion tokens using the Muon optimiser and FP8 mixed precision. It works with text only.

The model was trained with a context window of up to 128,000 tokens, but Baidu says it supports contexts of up to one million tokens for fine-tuning and inference. Weights are available in FP8 and BF16, and the card lists support for serving the model through the open-source SGLang and vLLM engines.

Baidu's own comparisons put the model against three Chinese base models of similar size: DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base, all run through the same evaluation pipeline. On the company's figures, ERNIE leads on most of the listed tests, scoring 93.3 on MMLU, 78.4 on MMLU-Pro, 60.2 on GPQA, 82.3 on HumanEval+ and 91.7 on the RepoQA long-context code test. Rivals come out ahead on several, including ChineseSimpleQA, BBH with chain-of-thought prompting, FullStackBench and parts of the HELMET and InfiniteBench long-context suites. The results are Baidu's own and have not been independently checked.

The model card links to three arXiv papers on techniques used in the model, covering Virtual Head Attention, a Multi-Router Latent mixture-of-experts design and a method called ECHO.

The weights come under Baidu's own ERNIE 5.5 Flash Base Model License Agreement rather than a standard open-source licence. Baidu released its ERNIE 4.5 family last year under Apache 2.0, so companies weighing commercial use will need to read the new terms.

The Decoder reported that a post-trained version is supposed to follow. The model card itself does not announce one or give a date.