Baidu's ERNIE 5.5 Flash Base: 280B parameters, license excludes the EU
Baidu released ERNIE-5.5-Flash-Base on 5 November: a 280B mixture-of-experts base model with 14B active and a 1M context. Its license excludes the EU, the UK and South Korea.
Baidu released ERNIE 5.5 Flash Base on 5 November 2026. It is the pre-trained checkpoint of ERNIE 5.5 Flash, a text-only mixture-of-experts model with 280 billion parameters and 14 billion active per token, under a license that does not apply in the European Union, the United Kingdom or South Korea.
What is new
This is a base model: the weights at the end of pre-training, before any mid- or post-training. It has no chat behavior, so you call it through a completions endpoint, and Baidu positions it for domain adaptation, instruction tuning and reinforcement-learning research. Baidu trained it from scratch on 33.5 trillion tokens with a training context of up to 128K, and the released configuration supports 1M tokens for fine-tuning and inference. It has 43 layers: one dense layer and 42 mixture-of-experts layers with 512 routed experts, 10 chosen per token, plus one shared expert.
In Baidu’s own table, the model is run through the same evaluation pipeline as the base models of DeepSeek-V4-Flash, Hy3-preview and MiMo-V2.5. It leads on most tests, for example 93.3 on MMLU against 89.4, 87.5 and 87.6, and 78.4 on MMLU-Pro against 69.2, 66.9 and 65.7. Rivals are ahead on a few, such as DeepSeek on ChineseSimpleQA and Hy3-preview on FullStackBench. All of these numbers are reported by Baidu.
Running it
Baidu publishes FP8 weights of about 288 GB and BF16 weights of about 573 GB. The FP8 files fit in the 640 GB of a server with eight 80 GB GPUs with room left for context, while the BF16 files leave little, so fine-tuning needs a multi-GPU server. The model card documents serving with SGLang and vLLM.
License
The weights use the ERNIE 5.5 Flash Base Model License Agreement v1.0, which is neither Apache nor MIT. It is valid only outside the EU, the UK and South Korea, and using, copying or distributing the model or its outputs in those places is not authorized. Non-commercial use is free. Commercial use is free while the related products or services earn up to US$50 million a year, if you show the exact words “Powered by ERNIE” and send Baidu a one-time notice. Above that you need Baidu’s written authorization before you cross the threshold, and the revenue of the whole product counts even if ERNIE is only a part of it.
The license also bars you from using the model or its outputs to improve any AI model other than the model itself or its derivatives, and the agreement is governed by Hong Kong law.
Why it matters
ERNIE 5.5 Flash Base is useful for researchers who want to run their own post-training. The license is the catch. Tencent’s Hunyuan Community License, used by HunyuanOCR, carves out the same three places, which is called a territory restriction, while other large Chinese models such as Hy4 preview and DeepSeek V4 have no such limit. If you work in Europe or Korea, you cannot use ERNIE 5.5 Flash Base at all, whatever your size.
AI licenses explained shows how to read clauses like these, and the commercial use list sorts models by what their licenses allow.
Sources
- ERNIE 5.5 Flash Base (Baidu ERNIE blog, 5 Nov 2026) ernie.baidu.com
- ERNIE-5.5-Flash-Base model card (Hugging Face) huggingface.co
- ERNIE 5.5 Flash Base Model License Agreement v1.0 huggingface.co
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| ERNIE 5.5 Flash Base Baidu | 280B 14B active | 1M | ERNIE 5.5 Flash Base (Conditional) | 193 GB Big-memory desktop | |
| Hy4 preview Tencent | 770B 49B active | 1M | Apache 2.0 (Permissive) | 531 GB GPU server | |
| DeepSeek V4 DeepSeek | 284B – 1.6T 49B active | 1M | MIT (Permissive) | 176 GB Big-memory desktop | |
| HunyuanOCR Tencent | 1.12B | 128K | Tencent Hunyuan (Conditional) | 0.773 GB Laptop or phone |