MoWorld: Domestic NPU Achieves 50FPS Interactive World Model
机器之心 · wechat · 2026-07-07
Moxin Technology officially released MoWorld, claiming it as the world's first Flash World Model (an ultra-high frame-rate interactive world model). Built on a 14B parameter MoE architecture, it achieves up to 50FPS real-time inference on Huawei Ascend NPUs, targeting autonomous driving and robotics control scenarios. Compared to GPUs, full-stack NPU optimization reduces inference costs by 70%, marking the first complete pipeline from pre-training to real-time deployment on a domestic NPU, and supporting long-term interactive video world generation. The technical report has been released, with model weights and code open-sourcing soon, alongside public access to a domestic NPU super-node service.
Related event: MoWorld: Real-Time Interactive World Model on Domestic NPUs(2 posts)→
More from Infra
- 12 KV Cache Reduction Techniques Every AI Engineer Should Understand, Explained — blaizedsouza · 2026-09-11
- The shadow GPU capacity market is formalizing, with Meta selling excess compute to outside buyers — DavidLinthicum · 2026-09-11
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11