Baidu has posted the pre-training weights of ERNIE-5.5-Flash-Base to Hugging Face and ModelScope. It is a text-only sparse mixture-of-experts model with 280B total parameters and about 14B firing per token, trained on 33.5 trillion tokens.
What it is not is a chatbot. The release is the checkpoint at the end of pre-training, with no mid-training and no instruction tuning. Baidu says a post-trained ERNIE 5.5 Flash will follow but gives no date.
A big model that stays cheap to serve
Each of the 42 expert layers holds 512 routed experts, and ten of those plus one shared expert activate for any given token. Inference cost tracks the active slice, while memory tracks the whole thing. In FP8 that means roughly 280 GB of weights, which will not fit on a single 80 GB card.
Baidu built three pieces of new machinery around it. Multiple independent routers score experts jointly instead of relying on one. A low-rank step lets attention heads share information and is fused into the kernel at inference. A module called ECHO reuses the final layer to predict future tokens and double as its own draft model.
The scores and the caveats
Against DeepSeek-V4-Flash-Base on Baidu’s own evaluation, the new model takes MMLU-Pro 78.4 to 69.2 and HumanEval+ 82.3 to 69.5. It loses ground on ChineseSimpleQA, 73.0 against 75.3, and several long-context tests.
Every benchmark here is self-reported under a setup the model’s author chose. Nobody independent has run it yet.
The bigger open question is legal. The card points to the ERNIE 5.5 Flash Base Model License Agreement v1.0 rather than granting a plain commercial right, and until that reads clearly, most teams will wait.
