MoWorld 2.0: Chinese startup folds 10+ physics factors into one 4D world model at 70FPS on Ascend NPUs
量子位 · wechat · 2026-11-07
Moxin (MoCore), founded by Zhejiang University PhD student Chen Tianrun, released the MoWorld 2.0 general 4D world model at the Wuzhen World Internet Conference. Beyond spatial generation and real-time interaction, it integrates 10+ physical factors—gravity, elasticity, buoyancy, fluids, soft/rigid bodies, heat—into one model, simulating forces, collisions and environmental change. Key techniques include a physics Gaussian renderer and multi-expert prediction over physics, pixel and Gaussian signals; generated scenes export as 3D/4D Gaussian assets editable in game engines, film pipelines and robot simulation (e.g., generating a walkable ancient market from a painting).
After 1 year of Huawei collaboration, task partitioning across multiple Ascend NPUs, sparse-attention optimization and mixed-precision quantization push real-time inference from 50 to 70 FPS, with inference costs as low as 30% of comparable GPU setups. An on-device app, "Mo Xiao Chong," already shipped on new Huawei phones (600M-param edge model on Kirin). The company raised 2B RMB across 5 rounds this year at a near-10B valuation, positioning itself as a "Chinese Cosmos" physical-world foundation model for automakers and embodied AI.
More from Multimodal
- Local pipeline showcase: MiniMax H3 refMod generations from reference image to 4K upscale — edisson75 · 2026-11-13
- Gaussian splats may be the best way to capture how 3D displays like LKG Portrait look IRL — nptacek · 2026-11-13
- AI-generated 'not canon' clip of Schwarzenegger driving with Michael Jordan — alanskimp · 2026-11-13
- H3-Image API gotcha: server keeps billing on client timeout — set max_retries=0 — xiaohu · 2026-11-13
- MiniMax open-sources H3-Image: native 16MP output, up to 16 reference images — xiaohu · 2026-11-13
- Daily AI painting made with GPT-6 Images 2.6 shows painterly winter village — DeryaTR_ · 2026-11-13