vLLM Adds Day-One Support for Baidu’s ERNIE 5.5 Flash Base

What happened

vLLM announces day-one support for Baidu’s ERNIE 5.5 Flash Base, an open base model with 280B total parameters and 14B active per token. The post reports 82.3% HumanEval+ and 40.1% LiveCodeBench pass@1 in 1-shot evaluations, compared with 69.5% and 31.9% for DeepSeek-V4-Flash-Base under the same setup. It describes serving with Baidu’s vLLM-ERNIE plugin on vLLM 0.32.0 and links to the model weights and recipe.

Read next

• More changes in today’s briefingGoogle DeepMind Rolls Out Gemini Omni 1.2 Flash for Longer Video Generation

The story so far

Baidu has released ERNIE 5.5 Flash Base as an open base model, and vLLM added day-one serving support through Baidu’s vLLM-ERNIE plugin on vLLM 0.32.0. The 280-billion-parameter mixture-of-experts model activates 14 billion parameters per input; its weights are available on Hugging Face in FP8 and BF16. Baidu says the model was pretrained on 33.5 trillion tokens, supports up to 128K context during pretraining and up to 1 million tokens during fine-tuning and inference, and leads similarly sized models across several benchmarks. vLLM reported 82.3% HumanEval+ and 40.1% LiveCodeBench pass@1 in one-shot evaluations, compared with 69.5% and 31.9% for DeepSeek-V4-Flash-Base under the same setup. The release does not include fine-tuning or post-training; Baidu plans a post-trained version.