Baidu Open-Sources 280B ERNIE 5.5 Flash AI Model
Baidu has open-sourced ERNIE-5.5-Flash-Base, a 280-billion-parameter AI model with 14B active per token, 1M-token context and a stricter custom license.
Baidu has open-sourced the weights of ERNIE-5.5-Flash-Base, a 280-billion-parameter AI model that activates only 14 billion parameters for each token it generates. The company released the model on November 5, with the weights available on Hugging Face in FP8 and BF16 formats. It is Baidu's first open release of a large model in about 16 months, since the ERNIE 4.5 series in June 2025, and the intervening ERNIE 5.0 and 5.1 models were never open-sourced.
A big model built to run cheap
ERNIE 5.5 Flash Base is a sparse mixture-of-experts model. That means it keeps a huge pool of 280 billion parameters in total, but only a small slice, 14 billion, is active for any single token. The design keeps the cost of each token low while the model still draws on a much larger pool of knowledge. Baidu pre-trained the model from scratch on 33.5 trillion tokens. It was trained with a context window of up to 128,000 tokens, and it supports a context of up to 1 million tokens for fine-tuning and inference, which lets it work with very long documents and long-running agent tasks in a single pass. The model has 43 Transformer layers, a hidden size of 4096, and 64 attention heads. Its efficiency trick is a hybrid attention design. In 37 of the 43 layers, every 128 historical tokens are compressed into a single cache record, which cuts the memory and computing cost of long texts. The other six layers keep fine-grained retrieval of the original content, so the model does not lose track of details. The model already works with the open-source inference frameworks vLLM and SGLang.
The benchmark claims, and the catch
Baidu compared the model against three rivals in the same weight class, DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, running all four through the same evaluation pipeline. Baidu says ERNIE 5.5 Flash Base ranked first in 24 of 30 benchmark tests, covering knowledge, reasoning, math, code, multilingual tasks, and long context. On MMLU, a widely used knowledge test, it scored 93.3 in a five-shot setting, ahead of DeepSeek V4 Flash Base at 89.4. Two caveats matter. First, these are Baidu's own numbers, and no independent party has reproduced them yet. Second, the comparison models are not each vendor's latest release, so the win says more about this size class than about the frontier as a whole. There is also a bigger caveat about what was released. This is a base model, the raw checkpoint at the end of pre-training, with no mid-training or post-training. It has no built-in chat alignment, no assistant behavior, no refusal policy, and no tool-use format. Developers are expected to fine-tune it themselves for a specific domain, add instruction training, or run reinforcement learning on top of it. Baidu says a post-trained version of ERNIE 5.5 Flash will follow.
The license is the real story
The release is open, but the terms are tighter than Baidu's last open release. ERNIE 4.5 used the permissive Apache 2.0 license. ERNIE 5.5 Flash Base ships under a custom license, the ERNIE 5.5 Flash Base Model License Agreement v1.0. The custom license does not permit use in the European Union, the United Kingdom, or South Korea. It also says that products or services that generated more than 50 million dollars in revenue over the past 12 consecutive months need separate written authorization from Baidu before commercial use. Any team planning to build on the model should read the license terms before adopting it.
Why it matters
Open-weight models in the 280-billion-parameter class used to be rare, and a 1-million-token context at this scale is a serious tool for developers building long-horizon agents and document-heavy applications. Baidu is betting that efficiency, not raw size, is what developers want, and it is handing the community a strong starting point for further training. But the license draws a line around who can actually use it. The regional ban and the revenue cap mean the model is open for many researchers and smaller teams, and closed for large commercial users in Europe and other restricted regions unless Baidu signs off. The weights are free to download, but the freedom to use them comes with conditions.
Sources
- Baidu - https://ernie.baidu.com/zh/release/ernie-5-5-flash-base/ - The Decoder - https://the-decoder.com/baidu-releases-ernie-5-5-flash-base-with-280-billion-parameters-as-an-open-source-foundation-model/ - ChainCatcher - https://www.chaincatcher.com/en/article/2302012