World Model 2026: Kunlun Wanwei Nails the AI Industry Timing

Short‑form video users notice AI‑generated bugs, AI‑written songs feel lacking in "human feel," and robot demos still fail at household tasks, all pointing to a common problem: current AI does not truly understand the rules of the real world.

At the World Artificial Intelligence Conference (WAIC) on July 19, Kunlun Wanwei’s chairman and CEO Fang Han asserted a core claim: 2026 is the year of the world model . This claim is supported by a cross‑modal model matrix that bridges virtual and physical realms.

The matrix consists of three layers:

Physical‑world layer: the Riemann‑1.0 embodied action model extends spatio‑temporal understanding to robot control, creating a general‑purpose robot brain that generalises across bodies.

Base layer: Matrix‑Game 3.5 provides foundational cognition, tackling long‑sequence consistency and real‑time interaction challenges in world modeling.

Digital‑content layer: Mureka V9.5 and O3 bring that cognition to music creation, elevating AI music from a mere generation tool to a compositional paradigm.

From virtual to physical, from base to scenario, a complete world‑model deployment system has taken shape.

Riemann‑1.0: A General‑Purpose Robot Brain

Riemann‑1.0 demonstrates the value of world‑model capability in the most complex physical domain. Traditional embodied‑intelligence demos fail when the environment changes (different desks, clothing, etc.) because the AI lacks physical rules and generalisation. Riemann‑1.0 breaks this deadlock, reliably performing ordered block stacking, complex kitchen tidying, and flexible garment folding.

Its architecture evolves from two prior generations:

First generation: specialised policy models trained end‑to‑end for a single robot, with poor generalisation and low training efficiency.

Second generation: VLA (vision‑language‑action) models that add large‑language‑model semantic understanding but still miss deep physical dynamics.

Third generation (WAM – World Action Model): Riemann‑1.0 integrates action generation, environment state evolution, and multi‑view visual prediction within a causal generation framework, achieving unified world understanding, future prediction, and closed‑loop control.

Architecturally, Riemann‑1.0 builds a unified action‑video modelling system using a Wan2.2 latent‑space DiT backbone, enriched with action tokens, state tokens, body‑specific mapping modules, and an action‑prediction head, enabling adaptation to diverse robot morphologies, sensor setups, and continuous action spaces.

Data infrastructure aggregates three sources—first‑person human video, UMI/exoskeleton glove data, and robot trajectory data—totaling 232,000 hours, addressing scale, generalisation, and supervision precision. A six‑stage automatic data‑cleaning engine validates the feasibility of "massive human‑video pre‑training + limited robot‑data fine‑tuning".

Training follows a three‑stage progressive strategy: LAM pseudo‑action pre‑training → real‑action supervised training → robot‑policy reinforcement training.

Benchmark results show clear superiority: LIBERO success rate 99.0% (industry top tier), RoboCasa‑365 success 62.6% (+8.4% over previous SOTA), RoboTwin2.0 success 94.3%. Ablation experiments reveal that adding human first‑person pre‑training improves long‑range seen‑scene success by 28.15% and unseen‑scene success by 16.97% compared to the same amount of robot demonstration data, surpassing the EgoVLA baseline by >25%.

Matrix‑Game 3.5: Real‑Time Interactive World Modeling

Matrix‑Game 3.5 addresses the long‑standing bottleneck where video models generate only short coherent clips; scene switches cause spatial drift, object penetration, and memory loss. The new model introduces three core innovations:

Patch Memory: history images are split into local patches, depth‑projected, and stored in 3D world coordinates. During generation, visible patches are retrieved and re‑projected to the current view, forming a spatial‑memory canvas.

PRoPE (Relative Pose Encoding): camera projection matrices are directly injected into the Transformer’s positional encoding, making geometric relations a natural part of attention.

Dynamic‑Static Dual‑Branch Memory: static scene patches are stored in Patch Memory, while dynamic entities are represented by separate reference tokens, preventing ghosting and misalignment.

These mechanisms eliminate the need for extra parameters while preserving the base video model’s generation quality and allowing rapid adaptation to different video backbones.

Matrix‑Game 3.5 also introduces a six‑stage automated data pipeline—camera‑pose & metric‑depth annotation, video‑prompt auto‑labeling, and a multi‑dimensional Scene‑Quality evaluator—forming a closed‑loop that solves the industry pain points of high data cost, annotation difficulty, and insufficient scale.

Engineered for efficiency, the 5‑billion‑parameter model runs at ~20 FPS in 720p on a single GPU. Its open‑source repository (https://github.com/Riemann-Dynamics/Matrix-Game-3.5) positions it as a "open‑source, universal base" contrasted with most WAIC submissions that are closed‑source or vertical‑only.

Mureka V9.5 & O3: AI Music with Reflection

AI music suffers from lacking emotional logic and human‑like phrasing. The Mureka series, champion of the global benchmark earlier this year, releases V9.5 and O3 to address this.

V9.5 builds on the MusiCoT (Music Chain‑of‑Thought) system, where the model first constructs a full‑song mental plan before generating audio, allowing structured composition, dynamic phrasing, and purposeful silence. Evaluation on 100 same‑scale test songs shows human‑voice quality good‑product rate 61.0% (+8% over previous), naturalness and expression closer to real singers, prompt compliance 97.0%, style adherence 95.7%, and overall usable rate 28.0%—the best among contemporaries despite a restrained arrangement density.

O3 adds test‑time scaling, giving the model a "reflection" ability: during generation it continuously revisits the original song goal, correcting deviations. This raises music‑quality good‑product rate from 35% to 43%, human‑voice quality from 60% to 68%, and overall usable rate from 35% to 39%.

Both versions were publicly demoed with prompt examples (soft piano, Chinese pop, atmospheric synths, progressive strings for V9.5; Dream‑Pop, reverb guitar, floating male vocals for O3). Notably, actor Huang Xiaoming publicly praised Mureka’s output, indicating entertainment‑industry acceptance of domestic AI music models.

Industry Context and Impact

WAIC featured over 1,100 exhibitors and 300+ global‑first products, making the AI industry’s trajectory clear: compute clusters are scaling (Huawei’s 8,192‑card Ascend node), robot hardware is diversifying (heavy‑duty humanoids, deformable robots), and large‑model parameter efficiency is accelerating (Kimi K3’s 2.8 trillion‑parameter open model narrows the gap with the world leader to weeks).

Matrix‑Game 3.5, Riemann‑1.0, and Mureka form a unified multimodal matrix that enables cross‑domain reuse: game studios can generate interactive worlds, digital‑twin platforms can build real‑time simulations, and embodied‑intelligence firms can train robot policies in a shared environment.

Open‑source releases (Matrix‑Game 3.5, DiT‑based video models) have already been adopted as baselines by industry leaders such as NVIDIA’s SANA‑WM and Adobe’s RELIC, and academic projects like Solaris (a multi‑person video world model) and Light Interaction (NVIDIA + Zhejiang University). This external validation underscores the technical value of the open‑source stack.

Conclusion

Looking back from the 2026 WAIC node, the year is likely to be marked as a watershed for world‑model development. Kunlun Wanwei’s declaration of a "world‑model year" and its rollout of a unified cognition foundation that spans digital creation, physical interaction, and artistic expression are accelerating the AI industry toward the next generational battlefield.

When AI can generate temporally consistent open worlds, understand musical emotion and structure, and perform generalized complex tasks in the physical world, the era of world models will truly have begun—mirroring the transformative impact of mobile internet and large‑language models.