Kunlun Tech Declares 2026 the 'Year of World Models' and Ships Matrix-Game 3.5

Core Highlights

Kunlun Tech chairman Fang Han announced at WAIC that 2026 is the 'first year of world models,' and in one go released the Matrix-Game 3.5 world model along with the Mureka v9.5 and O3 music models. The most watched is Matrix-Game 3.5: a world model capable of generating interactive game footage in real time, pushing 'letting AI build worlds' from demo videos into a usable tool, and giving the repeatedly discussed concept of 'world model' its first deployable open-source Chinese implementation that local developers can actually run and modify. Calling a year the 'first year' of a field is a bold marketing claim, but backing it with a runnable, open architecture is what separates a slogan from a real platform bet on where the industry is heading.

What the Capability Actually Does

The core breakthrough of Matrix-Game 3.5 is patch-level memory injection. Simply put, while generating each frame, the model can continuously write key scene information into memory at the patch level, so that across long sequences the objects, positions, and logic in the picture stay consistent, avoiding the kind of collapse where something is on the table one second and gone the next. On performance, the 5B-parameter version reaches 20FPS real-time generation on a single graphics card at 720p resolution, already meeting the basic requirements of an interactive experience. Also released at the same event were the Mureka v9.5 and O3 music models, completing Kunlun Tech's map of AIGC content offerings and extending the company's reach from vision all the way to audio, so a creator can pair generated worlds with generated soundtracks in one ecosystem. The world model and the music models together hint at a vertically integrated content studio, where the visual stage and the audio score are both machine-generated and mutually consistent.

Technical Details

Patch-level memory injection is a more stable scheme than traditional inter-frame propagation: instead of simply feeding the previous frame back, it injects structured memory into the generation process, supporting long, controllable evolution of world state. The single-card 20FPS depends on efficient diffusion inference and memory scheduling, letting consumer-grade hardware also run a world model without piling up a cluster of compute that only a hyperscaler could afford. This idea of explicitly managing state is also what fundamentally distinguishes a world model from an ordinary video-generation model that merely produces pretty clips without any coherent internal notion of a persistent place. Because the memory is structured rather than just pixels, the system can reason about what changed and why, which is exactly what an interactive user expects when they nudge the world and want a sensible reaction.

Comparison with Competing Approaches

Compared with overseas world models that mostly stay at research demos, Matrix-Game 3.5 emphasizes 'real-time, controllable, deployable,' and its core architecture is already open-source, so developers can take it, modify it, and integrate it. This makes it more down-to-earth than purely closed solutions in directions like game prototyping, virtual production, and embodied simulation, and it is far easier for domestic teams to drop directly into their own product pipelines without negotiating a commercial license for every deployment. Where a closed model might show a stunning reel but forbid commercial use, the open architecture here is built to be bent to a team's own game or film, which is the difference between inspiration and infrastructure.

Industry Impact and Use Cases

Simply put, the world model is the next stop for generative AI moving from 'make one image, cut one clip' to 'generate an interactive world.' For indie game developers, film previsualization, and robot simulation training, this kind of open-source, real-time-runnable model means the era of building worlds at low cost is opening up, and real-time world generation that only big vendors could afford to play with is now something small and mid-sized teams can pick up as well, lowering the cost of experimentation dramatically. The deeper implication is that simulation itself becomes cheap: a robot or a game character can be dropped into a generated world and trained against it, turning world models into a kind of infinite, on-demand practice environment. For a game studio, that means prototyping a level, a weather system, or an NPC behavior by simply describing it, then watching the world render the consequences in real time instead of waiting on an art pipeline that costs weeks and headcount. For filmmakers, the same engine becomes a previsualization stage where a director can iterate on blocking and lighting before a single frame is shot for real, de-risking expensive production decisions. And because the architecture is open, none of this is locked behind a vendor's permission or a per-seat license that punishes a small team for scaling up, which is usually the quiet reason breakthrough demos from big labs never reach the people who would actually build products on top of them.