ERNIE 5.5 Flash Base launches: An open model with 280 billion parameters

Baidu released its open foundation model ERNIE 5.5 Flash Base on 5 November 2026. The model’s weights are available on Hugging Face in FP8 and BF16 formats. Baidu has released the model without fine-tuning or post-training, leaving developers to decide how to adapt and train it for their own needs. The company says a post-trained version is also on the way, but has not given a release date.

Model scale and training process

ERNIE 5.5 Flash Base uses a mixture-of-experts architecture. The model has 280 billion parameters in total, with 14 billion parameters activated for each input, according to the source. It was pretrained on 33.5 trillion tokens.

The context length during pretraining reached up to 128,000 tokens. Baidu says the model is designed to support context lengths of up to one million tokens during fine-tuning and inference. These figures refer to different stages: the first describes the pretraining context, while the second is the target capacity for later use.

The model was released as a foundation model, not as a ready-made product tailored to a specific application. Since it does not include fine-tuning or post-training, teams that want to carry out those steps will need to manage their own development process. Baidu has yet to announce when the post-trained version it says is coming will be released.

Weights and benchmark results

Making the weights available on Hugging Face in FP8 and BF16 gives teams evaluating the model two format options. However, the source provides no information about pricing, licensing terms or recommended system requirements. Access to the weights alone therefore does not clarify the terms for commercial use or whether the model will run on specific hardware.

In benchmarks cited by Baidu, the model competes with open foundation models of a similar scale on reasoning, coding and long-context tasks. The comparison also includes DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base. Baidu says the tests across exam, STEM, reasoning, knowledge, coding, multilingual and long-context categories used the same evaluation pipeline. The performance results are based on Baidu’s reporting.

Implications for architecture and visualization workflows

ERNIE 5.5 Flash Base is not presented in the source as a tool designed specifically for architecture or 3D software. It therefore cannot be said to integrate directly with these applications or offer a specific function such as image generation. Architecture firms and visualization teams developing their own AI-powered tools could evaluate the model for long-context or coding tasks; these are potential use cases inferred from the available information.

Before putting it into practice, teams should check hardware requirements, compatibility of the FP8 and BF16 formats with their existing environment, and licensing terms, given the model’s overall scale. The source does not specify a graphics card or minimum system requirements, so it is not possible to draw a definitive conclusion about suitable hardware.