Baidu releases ERNIE 5.5 Flash Preview
The first post-trained model on the 280B-parameter ERNIE 5.5 Flash base, with 14B active parameters and a 1M-token context; API access is live and Apache 2.0 weights are promised.
- Models & capabilities
- Minor
Baidu released ERNIE 5.5 Flash Preview, the first post-trained model built on the ERNIE 5.5 Flash base checkpoint it published six days earlier. Baidu’s page gives it 280 billion parameters in total, 14 billion active, and a context window of 1 million tokens. The API is available on Baidu’s Qianfan platform, and the page links a free listing on OpenRouter.
Baidu describes the model as an “intermediate checkpoint from post-training”, already competitive with models of its size, and says some capabilities “may still fall short” while post-training continues. It emphasises coding, 3D modelling, knowledge work and content creation, and says its training environments include 300,000 knowledge-work workspaces and more than 150,000 video-creation environments. Showcase outputs include a game built over an 11-hour coding task and a short animated film; Baidu notes the model is text-only and could not see the game screen, and that some footage in a video example came from a separate model, Seedance 2.5.
Benchmark results appear on the page only as images, with no numeric scores in the text, and they are Baidu’s own; no independent evaluation was cited.
On weights, the page says they “will be open-sourced under Apache 2.0 soon”, with no date. That is a promise, not a release, and it differs from the base model’s custom licence, which excludes the EU, the UK and South Korea.