DeepSeek V4 Pro 0813 is now available via API only through OpenRouter, as the company has not published a dedicated announcement page. The model appears likely to release open weights soon, following the pattern of earlier versions from April and July which are hosted on Hugging Face.
The most notable change is how the model renders images based on reasoning levels. Low settings produce flat vector illustrations, medium settings generate loose outlined cartoons, and high settings create detailed scenes with specific colour palettes. This consistency across reasoning tiers is unusual compared to other models that typically ignore such parameters. Benchmark data remains scarce because official results were shared in a WeChat group, then copied to Reddit and deleted before appearing as ASCII art on Hacker News.
* Low reasoning yields flat vector illustrations
* Medium reasoning produces loose outlined cartoons
* High reasoning generates detailed colour scenes
