Baidu Opens ERNIE 5.5 Flash Weights, but Bars the EU, UK and South Korea

Baidu just cracked open its vault again. The Chinese tech giant published the model weights for ERNIE-5.5-Flash-Base, its first open-source release since the ERNIE 4.5 series landed in June 2025 — a gap of about 16 months. The models in between, ERNIE 5.0 and 5.1, stayed locked up.

This one is built for speed over brute force. Wenxin 5.5 Flash carries 280 billion parameters in total but wakes up only 14 billion of them per generated token. It was pre-trained from scratch on 33.5 trillion tokens and can handle up to 1 million tokens of inference context.

Baidu also reworked how the model chews through long text. Inside its 43-layer network, 37 layers squash every 128 historical tokens into a single cached entry, while the last 6 layers keep a fine-grained grip on the original content. The point: less memory and compute burned on sprawling documents.

In a shared evaluation against DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base, Baidu's model finished first in 24 of 30 benchmark tests. It's a strong showing in a domestic field that keeps getting more crowded.

What developers actually get is the plain text Base model, with no pre-training or post-training baked in. Baidu says it's aimed at teams that want to fine-tune for their own domains, run instruction training or do reinforcement learning. A post-trained ERNIE 5.5 Flash is promised shortly.

The fine print is where things get sharp. ERNIE 4.5 shipped under Apache 2.0; this release swaps in a custom license that forbids use in the European Union, the United Kingdom and South Korea. Anyone whose product or service pulled in more than $50 million over the prior 12 consecutive months needs written authorization from Baidu before going commercial.