The 280B model activates 14B parameters per token, claims a 1M-token context and 200 tokens per second, and is free on OpenRouter ahead of its weights release.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [x.com](https://x.com/Baidu_Inc/status/2120391732248637734)
Why it matters #
Baidu is testing whether a 280B model with only 14B active parameters can deliver agent-grade coding and knowledge work at small-model inference cost. The November 25th Apache 2.0 weights release will let developers verify the speed, context, and benchmark claims outside Baidu's hosted API.
Baidu released ERNIE 5.5 Flash Preview on November 11th, 2026, positioning the model on inference efficiency: 280 billion total parameters, 14 billion active per token, a 1 million-token context window, and generation speed above 200 tokens per second, according to Baidu's announcement on X. Baidu said the model weights will be released by November 25th.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.