Kunlun Tech Declares 2026 the Year of World Models: Matrix-Game 3.5 Hits 20FPS on a Single GPU with Patch-Level Memory Injection

Source: Kunlun Tech / Economic Daily / NetEase Tech URL: https://mp.weixin.qq.com/s/LidvGePhOOoUY3KTor_wg Date: 2026-07-19 (WAIC 2026)

Event Overview

At WAIC 2026, Kunlun Tech hosted a dedicated forum titled "World Models and Multimodal Paradigm Shifts" at the West Bund International Convention and Exhibition Center, making several major announcements:

1. Fang Han formally declared "2026 the Year One of World Models" — arguing that after two years focused on generation (text, images, video, music), AI's core focus is shifting to "understanding + interaction." 2. Matrix-Game 3.5 released — an interactive world model featuring patch-level memory injection, real-time 720p generation at 20FPS on a single GPU with a 5B-parameter model, one-minute memory retention, and an open-source core architecture. 3. Mureka v9.5 + O3 music models — v9.5 upgraded its base model with instruction precision improved from 6.92 to 7.62; O3 uses test-time scaling to extend MusiCoT reasoning with hindsight bias correction and self-revision. 4. Riemann-1.0 robot (Riemann Dynamics, Prof. Liu Yang) — trained on 232,000 hours of multi-source embodied interaction data, achieving dual SOTA in simulation and real-robot benchmarks. 5. Keynote by Academician Zhou Zhihua — distribution-matching techniques compress rollout error from quadratic to linear growth, and an inverse Bellman equation reduces required sample size to 1/√T.

WAIC 2026 itself set records: exhibition area exceeded 100,000 square meters for the first time, with over 1,100 companies, 140+ forums, and 9 Turing/Newsela... Turing Award and Nobel laureates in attendance.

Technical Deep Dive

Matrix-Game 3.5: Patch-Level Memory Injection

Traditional world models use "Frame-level FOV Memory" — storing entire historical frames as memory. Problems: exploding compute, poor cross-frame consistency, and no editability.

Matrix-Game 3.5 splits historical frames into patch memory with 3D coordinates:

| Step | Action | |---|---| | 1. Lift historical frames | Use Depth + Pose to lift pixels into world coordinates | | 2. Retrieve for current frame | Retrieve patches visible in the current frame's camera frustum (FOV) | | 3. Reproject to view | Reproject visible patches into the current viewpoint as a mosaic | | 4. Inject into DiT | The mosaic enters the DiT backbone as Memory Tokens |

This upgrades "Frame-level FOV Memory" to "Patch-level FOV Memory," delivering four key advantages:

• More accurate: spatial memory retrieval + cross-frame consistency
• More stable: stable generation under camera motion
• More native: In-context Learning preserves the base model's dynamic generation ability
• Editable: users can freely edit memory blocks — the key to controllability

Performance: How 5B at 20FPS 720p Works

Inference acceleration relies on three techniques:

1. Reduced model throughput — cutting redundant per-token computation 2. DiT optimization — structural pruning + operator fusion 3. VAE pruning — lightweight decoding

The data pipeline: 5 million high-quality video clips + 10,000 training hours + 1,200 game scenes. The output format is a Video + Pose + Action + Language quadruple, directly feeding embodied world-action models.

Open-Source Ecosystem: Matrix-Game Is the Llama of World Models

From 1.0 to 3.5, Matrix-Game has remained open source. Key milestones:

| Version | Key open-source milestone | |---|---| | Matrix-Game 2.0 | First open-source real-time interactive world model | | Matrix-Game 3.0 | First open-source solution addressing the memory problem | | Matrix-Game 3.5 | Core architecture open-sourced |

Academia and industry already build on Matrix-Game:

• Prof. Saining Xie's team (NYU, author of DiT) built Solaris, the world's first multi-player video world model, on Matrix-Game 2.0
• NVIDIA + Zhejiang University's Light Interaction was developed on Matrix-Game 3.0
• NVIDIA SANA-WM and Adobe RELIC use Matrix-Game as a comparison benchmark

This is the closest a Chinese team has come to the de facto standard position in the open-source world-model ecosystem.

Mureka v9.5 + O3: Music Models Enter the "Versioned Production" Era

v9.5's core upgrade is instruction precision rising from 6.92 to 7.62, with improvements across melody, vocals, audio quality, and arrangement — marketed as "the AI music model with the least AI flavor."

O3 takes the test-time scaling + MusiCoT route: the point is not generating more or thinking longer, but letting the model continuously review its current expression — whether local melodies serve the whole piece, whether arrangement masks vocals, whether emotion is spent too early, whether structure drifts from original intent. Extra inference space is used for hindsight bias review and self-correction.

Mureka has evolved from a single generation tool into a versioned production platform: model training → inference-time selection → versioned creation tools → platform distribution. The Studio converts DAW operations into instruction-based creation.

Zhou Zhihua: Theoretical Breakthroughs for Decision-Level World Models

Zhou divides world models into perception, understanding, and decision levels. Decision-level models face two theoretical bottlenecks:

1. Compound error growing quadratically — after T steps, error is T² times single-step error, making long sequences infeasible. 2. Excessive sample complexity — requiring exponential training samples.

His team's breakthroughs:

• Distribution matching: compresses rollout error from quadratic to linear, making long-horizon rollout viable
• Inverse Bellman equation: reduces required samples to 1/√T

These provide a theoretical basis for decision world models in high-cost real-world tasks like fighter-jet control and smart factories — moving world models from "games/video" toward "physical decision-making." He also proposed "learnware": models paired with AI-generated specifications, enabling heterogeneous models to be reused and assembled without exposing data.

Fang Han's Three Industry Predictions

1. World models move beyond screens into the physical world — embodied intelligence requires robots to simulate environments before acting; competition shifts to general-purpose capability in unfamiliar environments. 2. Gaming will be transformed first — open-world games no longer need years of manual construction by hundreds of people; Matrix-Game 3.5 lets virtual worlds grow in real time with player actions. World models will become gaming infrastructure within three to five years. 3. AI music and video become mass creative tools — Mureka, China's strongest and a global top-two music generation model, lowers creation barriers.

Why This Matters

1. A Chinese team holds the de facto open-source standard for world models — Xie's Solaris, NVIDIA + Zhejiang's Light Interaction, plus NVIDIA SANA-WM and Adobe RELIC benchmarks confirm it. 2. 5B at 20FPS 720p on a single card is an engineering feat — game developers, creators, and indie studios can actually run it. 3. Patch-level memory injection addresses world models' core problem — memory is the essential difference between world models and generative models; 3.5 engineering it to be "editable" in open source is a key H2 2026 technical inflection point. 4. Decision-level theory breakthroughs — Zhou's work mathematically solves sample complexity, a rare case of theory and engineering advancing together. 5. WAIC 2026 scale — 1,100+ companies, 3,000+ exhibits, 300+ global product debuts signal China's upgraded home-field AI event.

Risks and Open Questions

• 5B parameters limit world knowledge — the model can understand that "balls bounce," but complex social systems like market crashes are beyond scope. Current world models suit physical-world modeling, not human society.
• Open source ≠ ecosystem — Llama took two years to build an ecosystem; Matrix-Game needs docs, community, toolchain, and plugins.
• The real gap with closed models is unknown — Google DeepMind and OpenAI's world-model progress is not public; "Year One of World Models" is more accurately "Year One of Open-Source World Models."
• Editable patch memory is a double-edged sword — user-editable memory requires redesigned safety review and content compliance.

One-Sentence Summary

5B at 20FPS on a single GPU + patch-level memory injection + open core architecture + academician-level theory breakthroughs — Kunlun Tech has claimed the first de facto standard position for Chinese AI in the open-source world-model ecosystem.

Source links:

• Original WeChat article: 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