AI Tools ↗

Kunlun Open-Sources Matrix-Game 4.0 for Real-Time Video-Model Gameplay

Summary

Kunlun Tech has released and open-sourced Matrix-Game 4.0, a system that combines coded game logic, an engine-maintained world state and a video model that renders the resulting scene. The company says it runs at 480p and 60 frames per second on one Nvidia RTX 5090, with under 150 milliseconds from input to response and peak memory use of 18.5 GiB. A coding agent can alter scenes and interactions from natural-language requests, while structural images and semantic conditions guide the renderer. The source links code, model weights and a project page, but its performance comparisons and claim of being the first consumer-GPU video-model game come from the developer and are not independently tested in the article.

Hana’s perspective

I find the architecture more important than the “first” claim. Keeping rules and state in code while using a video model as the renderer addresses a basic weakness of purely generated worlds: attractive frames are not enough if actions have no persistent consequences. A 5090 is expensive, but local real-time operation is still a meaningful reduction from data-center hardware. The open release gives developers a chance to test whether the reported speed survives new scenes and whether visual consistency is good enough for production. Those independent tests matter more than the launch comparisons.

Are you a writer with a great story? We'd love to hear from you: APPLY HERE