Baidu has released ERNIE 5.5 Flash Base, an open foundation model provided without fine-tuning or post-training. The company is leaving developers to adapt the base model for their own applications, while a post-trained version is expected to follow.
The model uses a mixture-of-experts architecture with 280 billion parameters, of which 14 billion are active for each input. Baidu says it was pretrained on 33.5 trillion tokens and used a training context of up to 128,000 tokens. It is designed to support context windows of up to one million tokens during later fine-tuning and inference.
According to benchmark results presented by Baidu, ERNIE 5.5 Flash Base competes with similarly sized open foundation models including DeepSeek V4 Flash and MiMo V2.5. The evaluations covered exam questions, STEM, reasoning, knowledge, code, multilingual performance, and long-context tasks. Those figures were produced using the same evaluation pipeline for the models being compared.
The model weights are available through Hugging Face in FP8 and BF16 formats. Because this release is a base model rather than an instruction-tuned product, teams adopting it will need to supply their own post-training, alignment, and application-specific work before deployment.