Baidu has open-sourced ERNIE 5.5 Flash Base, a 280-billion-parameter language model it says outperforms rival open models from DeepSeek, Tencent and Xiaomi — its first open large model in about 16 months. The weights are available now on Hugging Face, in FP8 and BF16 formats, for researchers and developers to download and fine-tune.
The model is a sparse mixture-of-experts design: 280 billion parameters in total, with only 14 billion activated for each token. Baidu says it was pre-trained from scratch on 33.5 trillion tokens, using the Muon optimizer and FP8 mixed precision. Its training context runs to 128,000 tokens and extends to one million tokens for fine-tuning and inference, aimed at long documents and agent-style workloads.
Unlike a finished chatbot, ERNIE 5.5 Flash Base is deliberately raw. Baidu released the checkpoint exactly as it came out of pre-training, with no mid-training or post-training, so developers can adapt it for their own domains, instruction tuning and reinforcement-learning research. A post-trained version, ERNIE 5.5 Flash, is promised to follow. The release is supported by the open-source vLLM and SGLang serving communities, and Baidu has published FlashMask attention kernels for Nvidia's Hopper and Blackwell GPUs.
A base model, not a chatbot
The architecture lists 43 layers, a hidden size of 4,096 and 64 attention heads. All layers except the first use a mixture-of-experts block with 512 routed experts — 10 activated per token — plus one shared expert.
The release also marks a return to open weights for Baidu. The company open-sourced the ERNIE 4.5 series in June 2025, but the ERNIE 5.0 and 5.1 versions that followed were kept closed. ERNIE 5.5 Flash Base is the first large Baidu model to be opened up since then, landing a day after Tencent released its Hy4 open-weights model — a sign of how fast China's open-model race is moving.
Baidu says it leads its rivals — on its own tests
Baidu compared the model against three publicly released rivals — DeepSeek's V4 Flash Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base — using the same evaluation harness, prompts and few-shot examples. On its own figures, ERNIE 5.5 Flash Base leads overall on knowledge, reasoning, maths, code, multilingual and long-context benchmarks, coming first in 24 of 30 tests.
The gaps are widest in code. Baidu reports 82.3 on HumanEval+, against 69.5 for DeepSeek's V4 Flash Base, and 40.1 on LiveCodeBench against 31.9. On MMLU, the five-shot knowledge test, it scores 93.3, ahead of DeepSeek's 89.4, Tencent's 87.5 and Xiaomi's 87.6. The numbers come with a caveat: they are Baidu's own, run on its own harness, with no independent reproduction yet, and rivals lead on a small number of individual tests.
The licence is the catch
The weights are open, but the licence is not the permissive Apache 2.0 Baidu used for ERNIE 4.5. The custom ERNIE 5.5 Flash Base licence excludes use in the European Union, the United Kingdom and South Korea, and commercial use above $50 million in related annual revenue needs Baidu's written authorisation. Smaller commercial users must give notice and carry a "Powered by ERNIE" attribution, and the licence bars using the model or its outputs to improve an unrelated AI system.
For developers outside those limits, the release offers a rare fixed starting point: a large, modern base model that can be fine-tuned and compared without repeating a 33.5-trillion-token training run. Whether it becomes a widely used foundation — or stays a research artefact boxed in by its licence — will depend on what the community builds on it, and on the post-trained version Baidu says is coming next.
