Baidu has open-sourced ERNIE 5.5 Flash Base, a 280-billion-parameter model built for research rather than ready-made chat. Announced by Baidu's ERNIE Team on November 5, 2026, with downloadable weights, it is one of the largest open models a Chinese lab has released — and one of the most interesting, because of what it is not.
What Baidu just released
ERNIE-5.5-Flash-Base is a text-only mixture-of-experts model. Of its 280 billion total parameters, only 14 billion are activated per token, so it runs far more cheaply than its size suggests. It was pre-trained on 33.5 trillion tokens using the Muon optimizer and FP8 mixed precision, with 43 layers and 64 attention heads. Training context is 128K tokens, while fine-tuning and inference support up to 1 million tokens. Downloads are available in FP8 and BF16 safetensors, and the model can be served with SGLang and vLLM.
The numbers that matter
On Baidu's own benchmark table, measured with the same harness against rival base models, ERNIE 5.5 Flash Base scores 93.3 on MMLU (five-shot), ahead of DeepSeek's V4 Flash Base at 89.4, Tencent's Hy3 preview at 87.5 and Xiaomi's MiMo V2.5 at 87.6. It also leads on HumanEval+ with 82.3 against DeepSeek's 69.5. But it is not a clean sweep: Tencent's Hy3 leads on FullStackBench and DeepSeek leads on InfiniteBench. These are Baidu's own claims, not independent results — but the MMLU gap is large enough to take seriously.
Why it's a base model, not a chatbot
This is a base model, and that distinction matters. The checkpoint ends at pre-training: it has not been tuned to follow instructions or chat. It is meant for researchers to fine-tune, adapt to specialist domains and test new training methods. A post-trained ERNIE 5.5 Flash will follow for everyday use. For now, the audience is labs and developers who want a strong, open starting point — not anyone looking for a finished assistant.
The license catch
Read the license before building on it. Baidu's custom license excludes the EU, UK and South Korea. Non-commercial use is free, while commercial use up to $50 million in annual revenue requires a one-time notice to Baidu plus prominent "Powered by ERNIE" attribution. Above $50 million, you need Baidu's prior written authorization. Outputs also may not be used to improve an unrelated AI or machine-learning system. Open weights, in other words, but not open season.
Why it matters
A 280B model with 93.3 MMLU, released as open weights, raises the bar for what researchers can fine-tune without a frontier lab's budget. The license limits who can use it commercially and where, so its real-world reach will depend on how those terms are enforced. But as a research artifact, it is a serious release — and a sign that China's open-model race is still accelerating.
Comments
Post a Comment