Matrix-Game 3.5WAIC first released! Open source world model brother, this time he wants to "create the world" and not just "shoot videos"

Intelligent Things Author | Wang Han Editor | Mo Ying If we want to talk about the hottest discussion direction in the field of AI in the past year, world model must have a place. In the past 18 months, world model companies have attracted more than US$10 billion in investment, and Turing Award winner Yann LeCun even predicted that it will replace large language models as the mainstream AI paradigm within three to five years. At the same time, the technical bottlenecks of the world model are being exposed. Due to weak long-term memory capabilities, insufficient camera control accuracy, and limited real-time interaction capabilities, most products remain at the stage of "generating videos" rather than "interacting the world." More importantly, world models on the market generally lack understanding of physical rules. For example, the model can generate a picture of a dog running, but it does not understand why the dog's four legs land in this order. Amid this chaos, the Skywork team of Kunlun Wanwei officially released the world model Matrix-Game 3.5 during the World Artificial Intelligence Conference WAIC 2026 on July 19. This model can continuously generate an interactive world based on user actions and event instructions, and maintain the consistency of scene structure, spatial layout, character identity, and dynamic content during the long-term generation process, providing a sustainable interactive generation basis for applications such as games, robots, embodied intelligence, and XR. 1. Upgrades in three aspects to achieve end-to-end real-time reasoning. Version 3.5 builds a complete system-level acceleration solution from three levels: model throughput optimization, DiT inference optimization and VAE decoding optimization to achieve end-to-end real-time reasoning. First of all, in terms of model throughput optimization, the Skywork team compressed the number of sampling steps to 3-step Sampling through distillation, and designed Chunked Causal Inference, which combined with KV Cache, Patch Latent and prediction Latent, to improve single inference throughput while maintaining generation quality, and significantly reduce the overall delay caused by multi-step sampling. Secondly, in terms of DiT inference optimization, Skyw