On November 11, Baidu released ERNIE 5.5 Flash Preview, the first post-trained, ready-to-use version of the ERNIE 5.5 Flash line: 280 billion total parameters with only 14 billion active per token, and a 1-million-token context window. It is no longer the pretraining-only checkpoint from November 5 that could only continue text — it is a complete model that can write code and call tools.
The self-reported scores published on Baidu's official page concentrate on agentic and coding scenarios: 46.5 on Terminal-Bench 4.0, 71.2 on DeepSWE v1.1, 61.9 on AutomationBench, and 36.4 on SpreadsheetBench V2 for office work. Baidu positions its strengths as coding, 3D modeling, knowledge work and content creation, with demos including generating a playable 3D game from a single prompt, a 3D model from a single image, working with WPS documents and spreadsheets, and generating video from a single prompt.
On access, the ERNIE 5.5 Flash Preview API is now live on Baidu's Qianfan platform; the model weights are not yet open, with Baidu saying they will be released under the Apache 2.0 license at an unspecified date. OpenRouter lists it at $0.28 per million input tokens and $1.13 per million output tokens, with cache reads at $0.028, hosted exclusively by Qianfan and not currently available through EU in-region routing.
First, Tell Preview and Base Apart
Two caveats matter. First, all of these benchmark scores are Baidu's own self-evaluations, with no independent third-party verification yet — leave room for doubt when comparing it with rival models. Second, Preview and Base are two states of the same architecture: Base is the pretraining checkpoint published on November 5 for teams to keep training themselves, while Preview is the finished, Baidu-trained product, currently available only through the API. Anyone hoping to deploy the weights directly will have to wait for the open release.
