Kunlun Tech Declares "2026 the Year of the World Model": 5B Model, 20FPS on a Single GPU, Patch-Level Memory Injection
*Source: Kunlun Tech / Economic Daily / NetEase Tech, July 19, 2026 (WAIC 2026)* *Original: https://mp.weixin.qq.com/s/LidvGePhOOoUY3KTor_w9g*
The Event
On July 19 during WAIC 2026, Kunlun Tech held a forum titled "World Models and the Multimodal Paradigm Shift" at the West Bund International Convention Center, announcing:
1. CEO Fang Han formally declared "2026 is the Year of the World Model" — after two years focused on generation (text, image, video, music), AI's core focus shifts to understanding + interaction. 2. Matrix-Game 3.5 release — an interactive world model with patch-level memory injection: a 5B model generating real-time 720p at 20FPS on a single GPU, with 1-minute memory capability and open-sourced core architecture. 3. Mureka v9.5 + O3 music models — v9.5 upgraded instruction accuracy from 6.92 to 7.62; O3 extends MusiCoT reasoning via test-time scaling with self-review and self-correction. 4. Riemann-1.0 robot (Prof. Liu Yang) — trained on 232,000 hours of multi-source embodied interaction data, achieving dual SOTA in simulation and real hardware. 5. Academician Zhou Zhihua's keynote — distribution matching compresses inference error from quadratic to linear; inverse Bellman equations reduce sample size to 1/√T.
WAIC 2026 itself set records: exhibition space over 100,000 m² for the first time, 1,100+ exhibiting companies, 140+ forums, and 9 Turing/Nobel laureates present.
Technical Core: Patch-Level Memory Injection
Traditional world models use Frame-level FOV Memory — entire historical frames as memory — suffering from compute explosion, poor cross-frame consistency, and non-editability. Matrix-Game 3.5 instead splits historical frames into Patch Memory with 3D coordinates:
| Step | Action | |---|---| | 1. Lift historical frames | Depth + Pose lift pixels into world coordinates | | 2. Retrieve for current frame | Retrieve patches visible in the current camera frustum (FOV) | | 3. Reproject | Reproject visible patches into the current view as a Mosaic | | 4. Inject into DiT | The Mosaic enters the DiT backbone as Memory Tokens |
Four advantages: more accurate (spatial retrieval + cross-frame consistency), more stable (camera motion), more native (in-context learning preserves base dynamic generation), and editable (users can freely edit memory blocks — the key to controllability).
Performance: How 5B at 20FPS 720p Works
Three inference acceleration techniques: 1. Reduced model throughput — less redundant computation per token 2. DiT optimization — structural pruning + operator fusion 3. VAE pruning — lightweight decoding
Data pipeline: 5 million high-quality video clips, 10,000 training hours, 1,200 game scenes. Output format is a Video + Pose + Action + Language quadruple, directly usable by embodied world-action models.
Open-Source Ecosystem: Matrix-Game as the "Llama of World Models"
| Version | Milestone | |---|---| | Matrix-Game 2.0 | First open-source real-time interactive world model | | Matrix-Game 3.0 | First open-source approach to the memory problem | | Matrix-Game 3.5 | Core architecture open-sourced |
Adoption:
- Saining Xie's team (NYU, DiT author) built Solaris, the first multi-player video world model, on Matrix-Game 2.0
- NVIDIA + Zhejiang University's Light Interaction is built on Matrix-Game 3.0
- NVIDIA SANA-WM and Adobe RELIC use Matrix-Game as a comparison benchmark
- Distribution matching: reduces inference error from quadratic to linear, making long-horizon rollout feasible
- Inverse Bellman equations: cuts required samples to 1/√T
- 5B parameters limit world knowledge — suited to physical-world modeling, not complex social systems.
- Open source ≠ ecosystem — building docs, community, toolchains takes years (as Llama showed).
- Gap with closed models — DeepMind/OpenAI progress is undisclosed; "year of the world model" is more accurately "year of the open-source world model."
- Editable memory is double-edged — safety review and content compliance need redesign.
- WeChat original: https://mp.weixin.qq.com/s/LidvGePhOOoUY3KTor_wg
- Economic Daily: http://city.ce.cn/fw/202607/t20260720_3096484.shtml
- China.com Tech: https://tech.china.com.cn/sx/20260720/413372.shtml
- Matrix-Game 3.5 GitHub: https://github.com/Riemann-Dynamics/Matrix-Game-3.5
This is the closest a Chinese team has come to a de facto standard in open-source world models.
Mureka v9.5 + O3: Music Enters "Versioned Production"
v9.5's key upgrade is instruction accuracy: 6.92 → 7.62, improving melody, vocals, audio quality, and arrangement — marketed as "the AI music model with the least AI flavor." O3 uses test-time scaling + MusiCoT, devoting extra inference to reviewing biases and self-correcting: whether local melodies serve the whole, whether arrangements mask vocals, whether emotion is spent too early. Mureka has evolved into a versioned production platform: training → inference-time selection → versioned creation tools (Studio converts DAW operations into instruction-based creation) → distribution.
Zhou Zhihua: Theoretical Breakthroughs for Decision-Layer World Models
Zhou divides world models into perception, understanding, and decision layers. Decision layers face two bottlenecks: (1) multi-step inference compound error grows quadratically (T² after T steps), and (2) sample complexity is too high. His team's solutions:
This provides a theoretical basis for decision world models in high-cost real tasks (fighter-jet control, smart factories). He also proposed learnware: models plus AI-generated specifications enabling heterogeneous models to be reused and assembled without sharing raw data.
Fang Han's Three Industry Predictions
1. World models move off-screen into the physical world — embodied AI requires robots to simulate environments before acting; competition shifts to general capabilities in unfamiliar environments. 2. Gaming changes first — open worlds no longer need hundreds of people over years; world models will become gaming infrastructure within 3–5 years. 3. AI music/video become mass creative tools — Mureka (China's strongest, top-two globally) lowers creation barriers.
Why It Matters
1. A Chinese team holds the open-source de facto standard in world models (Solaris, Light Interaction, NVIDIA, Adobe all use/benchmark Matrix-Game). 2. 5B at 20FPS 720p on a single GPU makes real-time world models accessible to indie studios. 3. Patch-level editable memory is a key H2 2026 technical inflection point. 4. Rare 2026 case of theory and engineering advancing in sync.
Risks and Open Questions
One-Line Summary
5B at 20FPS on one GPU + patch-level memory injection + open-sourced core architecture + academician-level theory: Kunlun Tech has claimed China's first de facto standard position in the open-source world-model ecosystem.