Baidu Releases ERNIE-5.5-Flash-Base Weights Under Restricted License

Baidu released weights for ERNIE-5.5-Flash-Base under a custom license that excludes use in the European Union, the U.K. and South Korea and adds a revenue-based approval rule for related products and services.

The model has 280 billion parameters and activates 14 billion for each generated token. Baidu pretrained it from scratch on 33.5 trillion tokens, and it supports context windows of up to 1 million tokens.

Its 43-layer design handles long inputs by compressing groups of 128 historical tokens into one cache record in 37 layers. The other six layers retain detailed access to the original content. Baidu designed the system to reduce GPU memory use and computational overhead during long-text processing.

The release is a text-only base model for developers to adapt with fine-tuning, instruction training and reinforcement learning. It has not undergone the additional training used to produce a post-trained version. Baidu said a post-trained ERNIE-5.5-Flash version will follow.

Baidu's evaluation compared the model with DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base. The company said ERNIE-5.5-Flash-Base ranked first in 24 of the 30 benchmarks.

The weights follow Baidu's June 2025 release of the ERNIE 4.5 series, a gap of about 16 months. Baidu did not release weights for ERNIE 5.0 or 5.1. The ERNIE 4.5 weights used the Apache 2.0 license; the new model has a custom license.

Under the new terms, commercial use of related products or services requires separate written authorization from Baidu if those products or services generated more than $50 million in revenue over the previous 12 months.