Baidu releases ERNIE-5.5-Flash-Base, a 280B-parameter open foundation model, on Hugging Face

Baidu has put the weights of ERNIE-5.5-Flash-Base on Hugging Face in FP8 and BF16 formats. It is a mixture-of-experts base model with 280B total parameters, of which 14B are active at a time. It was pre-trained on 33.5T tokens with a 128K training context, and Baidu says it can handle up to 1M tokens during fine-tuning and inference. It has had no fine-tuning or post-training, and a post-trained version is only described as expected later, with no date given.

Researchers and enterprises that want to build their own chat or coding models now have another large open base model from a major lab, which gives them more choice and some leverage over closed providers. The practical audience is narrow, though. A raw base model needs substantial post-training and serious hardware before it is useful, so most developers will wait for the post-trained release. The licence terms are not stated in the coverage and will decide how far commercial use goes.

The benchmark claims against DeepSeek V4 Flash and MiMo V2.5 are Baidu's own figures, with no independent replication, so treat "competitive" as unproven. The 1M-token figure is a design target for fine-tuning and inference, not what the model was trained on (128K), and long-context quality at that length is untested. The coverage links to no model card or benchmark source, so the specifications are checkable but not yet checked here.

1. 01Independent base-model benchmark results (for example community evaluations on reasoning, code and long-context tasks) against DeepSeek V4 Flash Base and MiMo V2.5 Base within the next few weeks.
2. 02The licence on the Hugging Face model card, and whether Baidu publishes a date for the post-trained version.
3. 03Download counts, community fine-tunes and quantised builds on Hugging Face, plus real reports of quality at context lengths well beyond 128K.