Baidu releases ERNIE 5.5 Flash base weights with a 1M-token context
The 280B-parameter sparse model activates 14B parameters per token, but its license excludes the EU, UK and South Korea.
Baidu has released FP8 and BF16 weights for ERNIE 5.5 Flash Base, a sparse mixture-of-experts model with 280 billion total parameters and 14 billion activated for each token. The company says it trained the model from scratch on 33.5 trillion tokens. It supports a 128,000-token training context and up to one million tokens for fine-tuning and inference.
This is a base checkpoint taken directly after pre-training, before instruction tuning or other post-training. That makes it a starting point for teams that want to build domain-specific models, compare alignment methods or run reinforcement-learning experiments—not a ready-made chat assistant. Baidu provides serving instructions for SGLang and vLLM.
In Baidu’s own unified evaluation, ERNIE 5.5 Flash Base scored 78.4 on MMLU-Pro and 82.3 on HumanEval+, ahead of the listed DeepSeek-V4-Flash-Base results. These are vendor-reported benchmark figures, and the model did not lead every test.
The weights also come with significant restrictions. The license does not grant use rights in the EU, UK or South Korea. Commercial use is free only when annual revenue from related products or services is no more than 50 million dollars, requires a one-time notice to Baidu and a visible “Powered by ERNIE” label; larger deployments need written authorization. Outputs may not be used to improve unrelated AI models.
Baidu
