DeepSeek-V4-Flash-Vision-Exp builds game worlds via screenshots

A developer using the DeepSeek-V4-Flash-Vision-Exp model reports it can create complete game worlds in about two days of work, according to a post on r/LocalLLaMA. The model's new vision capability lets it take game screenshots, which the developer says it uses to generate and correct game models and textures until they look right, fix visual artifacts and glitches, and write scripts that capture screenshot sequences for animations. The developer also credits the model with play-testing the game, including its UI and game mechanics, and says the results let them build a compelling game world in a couple of days. The work follows an earlier experiment in which the same developer said the Qwen3.8-Flash-Next model one-shotted a Cat-Hunt game demo. The full game produced with the DeepSeek model is linked in the post. The account is a single developer's report, and the underlying game and model behavior have not been independently verified.