Kunlun Tech Declares 2026 the Year of World Models, but Matrix-Game 3.5 Still Faces a Real-Time Test

Kunlun Tech declared 2026 the “year of world models” and introduced Matrix-Game 3.5 at the World Artificial Intelligence Conference in Shanghai. The company says its five-billion-parameter model can generate interactive 720p video at about 20 frames per second on one GPU. That claim places efficiency, rather than visual spectacle alone, at the center of the release.

Matrix-Game 3.5 arrived alongside the Mureka V9.5 music generator and Mureka O3 music reasoning model. Kunlun Tech also presented Riemann-1.0, a model designed to connect visual prediction with robotic actions. Together, these releases frame world modeling as a shared foundation for games, media, simulation, and robotics.

However, the available evidence supports only part of that broad story. Kunlun Tech has released Matrix-Game 3.5 code and base model weights, but its repository says the distilled real-time models are still coming. The immediate contest is therefore between Kunlun Tech’s open, efficient infrastructure strategy and the practical limits of its current release.

Google DeepMind’s Genie 3 already established a visible benchmark for navigable AI environments. Nvidia’s Cosmos platform is pushing world models toward robotics, autonomous vehicles, and synthetic training data. Kunlun Tech is trying to bridge those directions with one architecture, while making that architecture accessible to outside developers.

Kunlun Tech’s 2026 World Model Push Starts With Matrix-Game 3.5

The announcement matters because Kunlun Tech presented world models as infrastructure, not simply as more capable video generators.

A world model predicts how an environment changes after an action. Unlike a conventional video generator, it must respond to controls while preserving objects, geometry, and previous events. The system needs to remember what happened beyond the current frame.

Kunlun Tech chairman and chief executive Fang Han used the company’s WAIC event to name 2026 the “year of world models.” The phrase is a strategic claim, not an industry consensus. It argues that generative AI is moving from producing isolated media toward maintaining environments that users or machines can continuously explore.

Matrix-Game 3.5 is the technical centerpiece of that argument. Kunlun Tech describes it as a memory-augmented model for long-running, camera-controlled video generation. It supports both first-person navigation and third-person scenes built around a recurring character.

The company’s model launch details describe three central changes. These cover geometric memory, camera-aware positional encoding, and separate treatment for moving subjects. Each change addresses a different way that generated worlds lose coherence.

The first problem is spatial forgetting. A conventional video model sees a limited window of recent frames. When the camera leaves a building and later returns, the model can redraw that building differently. Doors move, signs change, and paths stop connecting.

The second problem is camera control. Video models can imitate motion without constructing a stable relationship between the camera and the scene. A requested turn can produce a plausible image transition while quietly breaking the environment’s geometry.

The third problem concerns moving objects. A memory system that stores everything can preserve an outdated position for a person or vehicle. Reusing that memory creates duplicates, visual trails, or “ghosts” when the object moves.

Kunlun Tech says Matrix-Game 3.5 addresses these failures while retaining a relatively small five-billion-parameter backbone. Its public repository identifies Wan2.2-TI2V-5B as the underlying text-and-image-to-video scaffold. First-person and third-person checkpoints are available under an Apache 2.0 license.

The company also claims roughly 20 frames-per-second generation at 720p on one GPU. That specification deserves careful wording because the current public package does not yet establish the complete result. Its documentation requires one Nvidia GPU with at least 40GB of memory for available base-model inference.

More importantly, the repository says the distilled real-time autoregressive models will be released later. The currently documented inference process uses 25 denoising steps by default. It produces blocks of 80 frames from an anchor image, camera path, and text prompt.

The release is therefore open in a meaningful but incomplete form. Developers can inspect the architecture, download base weights, and run supplied examples. They cannot yet reproduce the headline real-time configuration from the released package alone.

That distinction creates the article’s central tension. Kunlun Tech announced an efficient, reusable world-model foundation. Its code supports serious technical inspection, yet the most commercially relevant performance claim still awaits a reproducible public release.

Why Patch Memory Is the Main Matrix-Game 3.5 Bet

Matrix-Game 3.5 treats memory as a geometric retrieval problem instead of asking the model to preserve an entire world inside recent frames.

Its main mechanism is called Patch Memory. A patch is a small region taken from an earlier image. The system estimates that patch’s depth and combines it with the camera’s recorded position.

Matrix-Game 3.5 then projects the patch into a three-dimensional coordinate system. When the camera approaches the same area later, the system identifies historically visible patches and projects them into the new viewpoint. Those retrieved details become a memory canvas supplied to the generator.

This mechanism differs from storing whole frames. A full frame contains useful scenery, but it also locks that scenery to an old viewpoint. Patch-level storage lets the system retrieve smaller pieces that match the current camera view.

It also differs from building one permanent three-dimensional reconstruction. A global reconstruction can accumulate errors when depth estimates or camera positions drift. Patch Memory instead maintains a flexible collection of localized observations.

The approach resembles retrieval in a knowledge system. The generator does not need every historical observation in its active context. It needs the relevant observations, aligned with the current request and location.

That parallel matters beyond video generation. Long-running AI systems often fail when they either forget earlier information or load too much irrelevant history. Selective retrieval can preserve useful context without forcing every past detail through the model each time.

Developers interested in such context problems can compare this approach with a searchable knowledge base. The underlying data differs, but both systems depend on retrieving the right memory at the right moment.

Kunlun Tech combines Patch Memory with Warped PRoPE, a camera-aware positional encoding method. Positional encoding tells a transformer how tokens relate across time and space. Standard video encodings generally represent time plus two-dimensional image coordinates.

Warped PRoPE adds information derived from camera projection matrices. This gives the model a representation of rotation, translation, and perspective between views. Kunlun Tech says the method requires no additional backbone parameters.

The combination has a clear division of labor. Patch Memory supplies earlier visual evidence from the correct region. Warped PRoPE tells the attention system how that evidence should align with the current viewpoint.

The architecture also separates static scenery from dynamic subjects. Buildings, terrain, and fixed objects enter the geometric memory. People, vehicles, and other moving subjects receive separate multi-view reference tokens.

Motion detection and segmentation filter moving subjects from Patch Memory. Otherwise, an earlier image of a car could remain embedded in the street after the car moved. The reference-token branch instead carries the subject’s identity and appearance through the sequence.

The released Patch Memory code supports the basic architecture description. It includes memory reprojection components, depth estimation dependencies, first-person samples, and third-person workflows. The repository also provides separate checkpoints for both viewing modes.

These mechanics make Matrix-Game 3.5 more than a renamed video model. They directly address the state problem that separates an attractive clip from an explorable environment. The model attempts to remember where things belong, not merely what the last few frames looked like.

Yet geometry is not the same as physical understanding. A model can preserve a doorway while generating incorrect collisions, object behavior, or causal outcomes. Patch retrieval improves visual continuity without proving that the model has learned reliable physics.

The mechanism also depends on upstream estimates. Incorrect depth or camera information can place a remembered patch in the wrong location. Dynamic-object filtering can fail when motion is subtle, partly obscured, or shared with camera movement.

Those limitations do not erase the architectural value. They define what independent evaluation must measure. Useful tests should include repeated returns to earlier locations, moving subjects, camera reversals, occlusion, and long sequences with structural changes.

Kunlun Tech’s Efficiency Claim Pressures Closed World Models

The strategic challenge to Google DeepMind is not that Matrix-Game 3.5 clearly produces better worlds, but that Kunlun Tech wants comparable interaction on accessible hardware.

Google DeepMind announced Genie 3 in August 2025 as a general-purpose world model. Its interactive environments run at 720p and 24 frames per second. Google said the system could maintain consistency for several minutes.

That provides a useful reference point, but not a clean benchmark comparison. Google has not published Genie 3’s model size, hardware requirements, training data, or complete architecture. Different prompts, actions, latency definitions, and output pipelines further limit direct comparisons.

Kunlun Tech’s headline rate is slightly lower at roughly 20 frames per second. Its claimed deployment footprint is more revealing. A five-billion-parameter model operating on one GPU would make experimentation available to a broader group than a closed hosted system.

Matrix-Game 3.5 also offers access to its base checkpoints and implementation. Researchers can modify memory retrieval, substitute inputs, evaluate failure cases, or fine-tune components. Genie 3 remains a controlled product experience rather than an openly inspectable research foundation.

This does not make Matrix-Game 3.5 the stronger system. Genie 3 accepts text prompts for diverse navigable environments and has received broader public exposure through Project Genie. Kunlun Tech has not provided equivalent evidence for generalization across the same range.

The comparison instead reveals two competing routes. Google can combine large proprietary datasets, specialized infrastructure, and integrated consumer access. Kunlun Tech is betting that smaller models, downloadable components, and reproducible research will accelerate improvement outside one company.

The evolution of Matrix-Game supports that efficiency narrative. Matrix-Game 2.0 used a few-step autoregressive diffusion process to generate interactive video at 25 frames per second. Its training pipeline included about 1,200 hours of annotated Unreal Engine and game footage.

Matrix-Game 3.0 then focused on long-horizon memory and higher-resolution generation. Its technical paper reported up to 40 frames per second at 720p with a five-billion-parameter model. However, that result used a multi-GPU inference configuration rather than one GPU.

The apparent reduction from 40 frames per second in version 3.0 to 20 in version 3.5 is therefore not a simple regression. The hardware context changed, while the memory design became more complex. Version 3.5 also adds explicit viewpoint geometry and separate subject tracking.

Real-time performance remains more complicated than one frame-rate number. Input processing, depth estimation, memory retrieval, video decoding, and display latency all contribute to responsiveness. A system can report rapid generation while still feeling delayed during control changes.

Kunlun Tech’s release pressures closed systems only if outside developers can reproduce the complete pipeline. The final real-time checkpoint must sustain its rate during extended interaction, not just short sequences. It must also preserve quality as its memory collection grows.

Open publication can accelerate that verification. Researchers can test failure modes that a conference demonstration avoids. They can inspect whether remembered objects survive camera turns, whether dynamic subjects remain stable, and whether controls respond consistently.

The company’s position also depends on licensing and implementation clarity. Apache 2.0 covers the Matrix-Game 3.5 project, while several third-party components carry their own terms. Developers still need to review dependencies and the underlying Wan model before commercial deployment.

For game studios, the attraction is clear. A locally deployable world model could generate exploratory prototypes, interactive backgrounds, or simulation variations without sending every frame to a hosted service. However, current models do not replace deterministic game engines.

Traditional engines guarantee collision rules, scripted state, and repeatable outcomes. Generative world models instead predict plausible next frames. That flexibility supports rapid creation, but prediction errors remain unacceptable for many production mechanics.

The near-term opportunity lies in hybrid systems. A conventional engine can maintain authoritative state, while a world model supplies visuals, alternative scenes, or training variations. Matrix-Game 3.5’s camera control and memory design fit that experimental role better than a claim that it generates complete games.

Open Code Does Not Yet Prove a Deployable World Model

The biggest uncertainty is the gap between Kunlun Tech’s real-time announcement and the exact configuration outsiders can run today.

The public repository provides two five-billion-parameter base models. One handles first-person generation, while the other supports third-person scenes with optional reference images. Both rely on a Wan2.2 video-model foundation and an external metric-depth model.

The documented hardware requirement is substantial. Generation at 704 by 1280 pixels peaks near 40GB of GPU memory. The supporting machine also needs at least 64GB of system memory, according to the project documentation.

Those specifications still fit one professional GPU, which is notable for experimental world modeling. They do not make the model practical on common consumer laptops or mainstream gaming cards. “Single GPU” should not be interpreted as ordinary local hardware.

The larger qualification concerns speed. The published repository explicitly says distilled real-time autoregressive models will be released soon. That means the downloadable base checkpoints do not yet let researchers independently verify the announced 20-frame-per-second mode.

Distillation compresses a slower generation process into fewer steps. Matrix-Game 3.5 uses a staged approach that trains a causal generator to imitate a more expensive bidirectional diffusion model. The company says this process preserves camera controls and memory conditioning over longer rollouts.

Until those weights arrive, developers can study the memory system and produce examples. They cannot fairly confirm the whole deployment claim. This is a release sequencing issue rather than evidence that the stated result is false.

Evaluation presents another challenge. World models need measurements that go beyond image quality. A visually sharp sequence can still forget a room, ignore an action, move an object incorrectly, or create inconsistent geometry.

Long-horizon tests should measure whether a scene remains stable after the camera leaves and returns. Action tests should calculate whether requested movement reliably changes the predicted world. Dynamic tests should track the identity and location of subjects across occlusion.

Physical consistency requires a separate layer of testing. A system should predict plausible effects after contact, movement, and environmental change. Visual memory alone cannot establish causal reasoning.

Kunlun Tech’s demonstrations reportedly included open environments and first-person game scenes. These examples help explain the intended use, but curated demonstrations are not independent benchmarks. The company has not published enough third-party results for Matrix-Game 3.5 to establish a performance ranking.

The same caution applies to the larger product matrix. Mureka V9.5 reportedly improves vocal quality, prompt adherence, arrangement, and perceived realism. Mureka O3 applies additional inference-time computation to review and adjust musical decisions during generation.

Kunlun Tech reported internal evaluation percentages for these music models. It said Mureka V9.5 reached a 61 percent acceptable rate for vocal performance and 97 percent for prompt control. These figures come from the company’s evaluation process and should not be treated as independent market benchmarks.

The company said Mureka O3 raised its listening-quality acceptance rate from 35 percent to 43 percent. It also reported an increase in vocal acceptance from 60 percent to 68 percent. Without public test data and third-party replication, these numbers mainly describe Kunlun Tech’s chosen development targets.

The music releases broaden the announcement but do not validate the world-model thesis. Generating a coherent song and maintaining a navigable environment involve different state, latency, and evaluation problems. Their connection is strategic rather than technical proof of one unified system.

Riemann-1.0 creates a more direct test of the thesis because robotic control depends on predicting physical consequences. Kunlun Tech describes it as a world action model that jointly represents video, robot state, and actions.

The company reported results across robotics benchmarks and described training with 232,000 hours of mixed human and robot data. These are company-supplied claims. Independent reproduction will require model access, evaluation details, and comparable test conditions.

For enterprise buyers, the verification gap should shape procurement questions. Teams should request full latency measurements, supported GPUs, energy use, sequence duration, and failure rates. They should also distinguish an available base model from an announced optimized checkpoint.

Developers should test their own environments rather than relying on conference examples. A world model trained around games may perform differently on warehouses, streets, homes, or industrial processes. Camera control does not guarantee accurate domain behavior.

The release nevertheless establishes a concrete base for experimentation. The code is public, the architecture is inspectable, and checkpoints are available. That is more useful than a demonstration with no technical access.

The careful conclusion is narrow. Kunlun Tech has released an open geometric-memory system for long-horizon video generation. It has not yet supplied everything needed to confirm its single-GPU real-time claim independently.

The World Model Race Is Splitting Into Games and Physical AI

Kunlun Tech’s larger bet is that one memory-centered foundation can serve interactive media and machine training, while rivals specialize around different deployment markets.

Google DeepMind presents world models as environments for agents and users. Genie 3 generates worlds ahead of the camera as someone navigates. Its value lies in broad creation, interaction, and potentially unlimited training environments.

Nvidia approaches the category through physical AI. Its world model platform combines generation models, data-curation tools, evaluation systems, and post-training workflows. Cosmos targets robots, autonomous vehicles, and smart infrastructure.

Cosmos 3, released in 2026, extends that stack with models and workflows for perception, prediction, and policy development. Nvidia’s advantage comes from connecting world models to simulation software, accelerated computing, robotics frameworks, and deployment hardware.

Kunlun Tech sits between these positions. Matrix-Game began with game-like environments and interactive video. Riemann-1.0 attempts to transfer world prediction into robot actions. Mureka and the company’s video models add media-generation applications around that core.

This middle position creates opportunity and risk. Shared models can reuse representation learning, data infrastructure, and video-generation research. However, games, music, and robotics do not share identical accuracy requirements or commercial channels.

A game prototype can tolerate visual improvisation. A warehouse robot cannot tolerate an imagined obstacle location. A music creator evaluates emotional and aesthetic quality, while a robot developer measures task success and safety.

The phrase “world model” can obscure these differences. Some systems predict future sensor observations. Others generate playable video streams. Others learn compressed internal states that help an agent choose actions.

Matrix-Game 3.5 belongs primarily to the interactive video branch. Its inputs include an anchor image, camera trajectory, and text prompt. Its output is a generated video sequence with a visualized memory state.

That can support synthetic environments, but it does not automatically provide a symbolic or physically authoritative simulation. It predicts what a scene should look like as the camera moves. Developers still need mechanisms for actions, objects, collisions, and measurable state.

The strongest near-term use case is rapid environment variation. A studio could explore scene concepts from an image and camera path. A robotics team could generate visual conditions for perception experiments, subject to domain validation.

Digital-twin developers could use the system to create alternative visual trajectories. They would still need verified geometry and sensor models for engineering decisions. A plausible image cannot replace calibrated simulation data.

The open model route gives Kunlun Tech a way to build influence before a mature commercial category exists. Researchers can adapt the system, publish extensions, and make Matrix-Game a common baseline. That visibility can attract developers even when direct revenue remains uncertain.

Earlier Matrix-Game versions have already appeared in follow-on research and comparative evaluations. The company points to academic projects that built multiplayer or efficient interaction systems on its earlier releases. Such adoption matters more than promotional rankings because it shows whether outside teams find the architecture useful.

However, Nvidia offers a formidable open alternative. Cosmos includes code, models, datasets, and evaluation tools under an open model license. Its ecosystem focuses directly on physical AI buyers with established hardware and simulation workflows.

Google applies different pressure through consumer distribution and model scale. Project Genie can expose interactive generation to users without asking them to configure checkpoints or professional GPUs. Accessibility through a hosted product can outweigh source access for many creators.

Kunlun Tech therefore needs more than a technically interesting release. It needs a reliable developer experience, maintained checkpoints, clear evaluation, and integrations with existing production systems. Open code starts an ecosystem, but sustained support keeps one alive.

The company must also show that its cross-domain strategy produces technical leverage. Shared video representations are useful, yet each product still needs specialized data, controls, and safety work. A portfolio of models is not automatically one coherent world-model platform.

For knowledge workers following these releases, source management is becoming part of the evaluation process. Claims, checkpoints, papers, and benchmark revisions can change quickly. A personal knowledge base can preserve those materials alongside later test results and deployment notes.

The industry’s direction is clearer than the eventual winner. Interactive world generation is moving from short demonstrations toward persistent memory, real-time control, and application-specific evaluation. Competition now concerns infrastructure and usability, not only video quality.

What to Watch After the Matrix-Game 3.5 Release

Three signals will determine whether Kunlun Tech’s “year of world models” becomes a technical milestone or remains an ambitious conference theme.

The first signal is the promised real-time checkpoint. Kunlun Tech needs to release the distilled autoregressive weights, exact hardware configuration, and complete latency methodology. Outside teams should be able to reproduce roughly 20 frames per second at 720p on one supported GPU.

Sustained performance matters more than a brief peak. Tests should run across minute-long sequences with memory retrieval active. They should measure generation speed, end-to-end control latency, GPU memory use, and quality changes as the sequence grows.

A successful release would strengthen Kunlun Tech’s efficiency argument against closed systems. A delayed or difficult-to-reproduce release would weaken the main differentiator. The base model alone cannot settle that question.

The second signal is independent long-horizon evaluation. Researchers need to test revisited locations, difficult camera paths, occlusion, moving subjects, and environmental changes. Results should separate scene recall from physical prediction.

A useful benchmark would ask the model to leave a location and return after many generated frames. It would then compare the reconstructed scene with the earlier view. Another test would track whether a moving subject retains identity without leaving a duplicate in memory.

These evaluations should include Matrix-Game 3.0, Matrix-Game 3.5, and other available world models under comparable conditions. They should disclose hardware, input resolution, sequence length, and sampling settings. Without common conditions, frame-rate and quality claims remain difficult to interpret.

Strong independent results would show that Patch Memory produces measurable gains beyond curated examples. Weak results would suggest that geometric retrieval solves only a limited part of long-horizon consistency.

The third signal is outside adoption. The most persuasive evidence will come from developers building applications that the original team did not script. Game prototypes, simulation tools, robotics datasets, and research extensions can reveal whether the model works as infrastructure.

Adoption should be judged by more than repository stars. Maintained integrations, published experiments, fine-tuned checkpoints, and repeated use in peer-reviewed work offer stronger signals. Commercial pilots with disclosed constraints would add another layer of evidence.

The adjacent releases also deserve observation, but they should not distract from this sequence. Mureka V9.5 and O3 need independent listening tests and creator adoption. Riemann-1.0 needs reproducible robotics results and clear access terms.

Kunlun Tech declares 2026 the year of world models because it sees memory, prediction, and interaction becoming one common AI layer. Matrix-Game 3.5 gives that claim a concrete architecture and an open repository.

The unresolved issue is execution. Can the released system sustain a coherent world, respond at interactive speed, and run on the stated hardware outside Kunlun Tech’s demonstrations?

Developers should begin with the available base model, record reproducible failure cases, and compare results after the distilled checkpoint arrives. That process will reveal whether Patch Memory is a durable foundation or one more promising answer to a very difficult problem.