Baidu releases open-weight ERNIE 5.5 Flash Base model

Baidu ↘ 2.18% released ERNIE-5.5-Flash-Base on Thursday, November 5. It is an open-weight sparse mixture-of-experts language model with 280 billion total parameters, of which 14 billion are activated per token, according to the company's official release post. The weights are the checkpoint taken at the end of pre-training, before any mid-training or post-training. Baidu pitches it as a shared starting point for domain adaptation, instruction tuning and reinforcement-learning research. The company says a chat-ready version is still to come: "This base model is only a starting point. The post-trained ERNIE 5.5 Flash will follow," the ERNIE Team wrote.ernie.baidu.com

Built for efficiency

Baidu says it pre-trained the model from scratch on 33.5 trillion tokens with a training context of up to 128,000 tokens, and that it supports a 1-million-token context for fine-tuning and inference. The model card lists 43 layers. Thirty-seven of them use a compressed attention scheme that stores one cache entry per 128-token block, and nearly every layer routes to 10 of 512 experts plus one shared expert, Unite.AI reported. Baidu said "efficiency was a design goal from the outset," and that it designed the model and its training system together.unite.ai+1

In Baidu's own tests, the model led rival base checkpoints DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base on most benchmarks. It scored 93.3 on MMLU and 82.3 on HumanEval+, but trailed on FullStackBench and InfiniteBench. No independent evaluation has been published so far.ernie.baidu.com+1

Availability and hardware

The weights are on Hugging Face in FP8 and BF16 safetensors formats. Developers can serve the model with vLLM and SGLang through the completions API, and Baidu has published a vLLM plugin that lists the post-trained ERNIE-5.5-Flash as "coming soon." Running it takes heavy hardware. The FP8 checkpoint alone is about 288 GB, and NVIDIA's ↘ 8.58% full-model fine-tuning recipe uses 16 nodes with four GB300 GPUs each.huggingface.co+3

A restrictive license

The model is released under Baidu's custom ERNIE 5.5 Flash Base Model License Agreement v1.0. Its rights apply only within a "Territory" that covers everywhere except the European Union, the United Kingdom and South Korea. Non-commercial use is free. Commercial users with up to $50 million in annual related revenue must send Baidu a one-time notice and display "Powered by ERNIE," while larger companies need Baidu's prior written authorization. The license also bars using the model or its outputs to improve unrelated AI systems, which limits distillation. For now, one analysis notes, "open" describes whether the weights can be downloaded rather than what users are permitted to do with them.huggingface.co+2