Baidu has released ERNIE-5.5-Flash-Base, a large open-weight language model, on Hugging Face. The model is a sparse mixture-of-experts design with 280 billion total parameters, of which 14 billion are active per token, and was trained on 33.5 trillion tokens.
The release is a base model, not a chatbot. Baidu's model card describes it as "the checkpoint at the end of pre-training, before any mid- or post-training," meaning it has not been instruction-tuned. It is text-only, was trained with a 128K context, and supports fine-tuning and inference at up to 1,048,576 tokens.
The architecture has 43 layers, one dense and 42 mixture-of-experts layers, each with 512 routed experts (top-10 routing) plus a shared expert. Weights are published in FP8 and BF16, and Baidu documents serving with SGLang and with vLLM 0.32.0 via a vLLM-ERNIE plugin. The model is released under Baidu's own ERNIE 5.5 Flash Base Model License Agreement v1.0 rather than a standard open-source licence.
On Baidu's own comparisons against other base models of similar size, ERNIE-5.5-Flash-Base scores 93.3 on MMLU, 78.4 on MMLU-Pro, 60.2 on GPQA, 82.3 on HumanEval+ and 40.1 on LiveCodeBench, ahead of DeepSeek-V4-Flash-Base on each of those. The scores are self-reported by Baidu.
For teams that want to do their own post-training, a strong base checkpoint at this size is more useful than another chat model, because it has not been shaped by someone else's alignment choices. The custom licence should be read before any commercial use.