DeepSeek V4 Pro 0813: A Release with No Announcement and Mysterious Pelicans

DeepSeek has released the latest version of its flagship model — DeepSeek V4 Pro 0813. The model is available only through the API, yet there is no coherent announcement page for it: Simon Willison, who reported on the newcomer, had to point to the OpenRouter catalog, since no official announcement could be found.

Will the weights be opened?

There is still no confirmation that DeepSeek plans to release open weights for the new model. But, as Willison notes, the indirect signs are encouraging: the weights of previous releases have already been published — both April's DeepSeek-V4-Pro and July's DeepSeek-V4-Flash-0731. Against that backdrop, opening the weights for the August version too looks likely.

Pelicans at three speeds

The author's most intriguing observation concerns Willison's signature test — the request to draw a pelican on a bicycle. The new model provides three levels of “reasoning”: low, medium and high. And the pelican came out noticeably different for each of them. According to Willison, he had never before seen such a spread between modes in any other model — usually the results differ far less.

The benchmarks took a rough road

The performance figures are a story of their own. Tracing their origin turned out to be almost a detective's task. Apparently the numbers were first published in DeepSeek's official WeChat group. From there they were copied into a Reddit post, which the moderators deleted as “low quality.” And only then did the data migrate to Hacker News — in the form of a table rendered in ASCII art.

Thus the global community has to judge the fresh model's characteristics from repeatedly forwarded text rather than from official documentation. For a release from a major AI lab this looks unusual, and it underscores once again just how differently Chinese and Western developers approach public announcements.

Source: Simon Willison