Back to news
launchBaidu2026-11-06

Baidu released ERNIE 5.5 Flash Base weights on Hugging Face on November 5 with 280B total and 14B active parameters

Baidu published the ERNIE-5.5-Flash-Base pretraining checkpoint on Hugging Face, a sparse mixture-of-experts model with 280B total and 14B active parameters pretrained on 33.5 trillion tokens, 128K training context and up to 1M for inference, in FP8 and BF16.

On November 5, Baidu released the model weights for ERNIE-5.5-Flash-Base on Hugging Face. This is a pure pretrained base model using a sparse mixture-of-experts architecture with 280 billion total parameters, activating 14 billion per request, pretrained from scratch on 33.5 trillion tokens with a 128K training context, and an architecture designed to support contexts up to 1 million tokens during fine-tuning and inference.

Unlike common chat models, the Base version has not undergone instruction tuning and cannot be connected directly to a chat interface or an agent, requiring teams to prepare their own data and compute budget for supervised fine-tuning and preference training. Baidu said a post-trained version will follow, and that it compared ERNIE-5.5-Flash-Base against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base using a single evaluation pipeline covering exams, STEM, reasoning, knowledge, code, multilingual performance and long context, claiming parity with comparable open base models on reasoning, code and long context tasks. These are Baidu's own results and no independent measurements are available.

Weights are provided in FP8 and BF16, with FP8 requiring significantly less memory for smaller-cluster experiments and BF16 being more convenient for fine-tuning. The license is Baidu's own ernie-5.5-flash-base-model-license-v1.0 rather than the Apache 2.0 used by most open models, so teams should review the terms before commercial use.

ERNIE 5.5 Flash Base百度开放权重混合专家Hugging Face