Baidu has released ERNIE-5.5-Flash-Base, an open foundation model that leaves fine-tuning and post-training to developers. That choice gives teams room to decide how to adapt the model for their needs, while also asking them to do more of the work themselves. Baidu says a post-trained version is planned.
A base model built for adaptation
ERNIE-5.5-Flash-Base is distributed as a foundation model rather than a finished assistant. It comes without fine-tuning or post-training, so developers receive a starting point and are expected to choose how to shape its behavior for their applications.
This approach makes the release a flexible but hands-on option. Teams interested in the model will need to decide what additional training or other adaptation their use case requires. The source does not describe those steps or specify when Baidu’s planned post-trained version will arrive.
Scale and context capacity
The model uses a mixture-of-experts architecture and has 280 billion parameters, with 14 billion active for each input. In this design, the total parameter count describes the model’s overall scale, while the active count indicates how many parameters are engaged for an individual input.
Baidu says ERNIE 5.5 was pre-trained on 33.5 trillion tokens, with a training context of up to 128,000 tokens. It is designed to support a context of up to one million tokens during fine-tuning and inference. That distinction matters: the stated pre-training context and the designed context capacity refer to different stages of working with the model.
Benchmarks and access
According to Baidu, the model competes in reasoning, code, and long-context benchmarks with similarly sized open foundation models, including DeepSeek V4 Flash and MiMo V2.5. These are claims attributed to Baidu; the source does not provide benchmark scores or further testing details.
The model weights are available on Hugging Face in FP8 and BF16 formats. Those formats provide two options for accessing the released weights, while the model’s base status leaves developers to determine how to prepare it for their own work.
What the release asks of developers
ERNIE-5.5-Flash-Base puts the emphasis on customization. Its scale, mixture-of-experts design, and long-context capacity describe what the model is built to support, but they do not make it a ready-to-use, post-trained system. Developers must decide whether the base release fits their goals and what further work is needed.
Baidu’s planned post-trained version may offer another way to use ERNIE 5.5, but the current release is deliberately a foundation. For now, the central choice is whether a team wants that room to adapt the model and is prepared to take responsibility for shaping it.