Matrix-Game 3.5 is an open-source multimodal world model from Kunlun Tech's Skywork team, built for developers and researchers who need real-time, controllable, interactive video worlds for games, robotics and embodied AI rather than one-shot clips. It is the latest step in Kunlun's 4+3 AGI strategy alongside its SkyReels video and Mureka music models.
Core Features
- A 5B-parameter distilled model that generates 720p video at roughly 20 FPS on a single GPU, with about one minute of memory.
- Patch-level memory injection: historical frames become 3D-coordinate Patch Memory, retrieved by camera frustum and re-projected as Memory Tokens into the DiT model for spatial consistency.
- PRoPE camera-pose relative encoding plus a dynamic-static decoupled memory that keeps static scene structure while tracking moving subjects.
- System-level acceleration (3-step sampling, Chunked Causal Inference, FFN INT8 quantization, about 75 percent VAE pruning) and new NPC interaction with multi-style switching.
Use Cases
- Game studios prototyping interactive environments without rendering every frame by hand.
- Robotics and embodied-AI labs using it as an infinite data engine for training.
- XR and simulation teams needing consistent, camera-controllable scene generation.
Pricing
Matrix-Game 3.5 is free and open-source under Kunlun's public release on GitHub (Riemann-Dynamics/Matrix-Game-3.5), with no paid plan. There is no paid tier, and it has become a benchmark for NVIDIA's SANA-WM and Adobe's RELIC world-model work.
Our Take
Best for researchers and game studios who want a real-time, open world model they can self-host. The trade-off is the learning curve and hardware: the 20 FPS figure is on a single capable GPU, and tuning memory blocks takes effort. Compare with other video models in our AI Video directory.
