No Access to MoWorld's Full Model, So I Tried Its On-Device Version

As someone who completely doesn't understand world models, I gave MoWorld a try.

Let me be clear first, I can't get my hands on MoWorld itself. It's made by Moxin Technology together with Zhejiang University's Academician Pan Yunhe team, running on Huawei Ascend NPUs. An NPU you can think of as a chip specifically for AI compute, same category as the GPU in a graphics card, but this set is domestic. Our lab doesn't have an Ascend machine, and applying doesn't get us a spot in line, so what I actually got hands-on with was its edge-side version.

Let me also explain world models. Most video generation models are: you give a sentence, it plays you a clip. A world model is different — it has to understand space; you move, and what comes next in the frame has to change accordingly, and change reasonably. There was a line in the launch materials I still remember:

The key difference between a world and a video is that a world is capable of real-time physical interaction.

Hands-on

The edge-side one is called MoWorld-2.0-flash, and it was used to make a HarmonyOS-exclusive app called MoXiaoChong, for doing 3D reconstruction of pets. I borrowed a classmate's HarmonyOS phone and installed it.

The process is roughly: take a few photos of the cat and upload them, the cloud side generates a 3D Gaussian model on the Ascend platform using a diffusion model, then sends it back to the phone for rendering. The term 3D Gaussian confused me the first time I saw it too — you can think of it as a 3D method that breaks an object into a bunch of little ellipsoids, more convenient than traditional meshes. The edge-side model is about 600 million parameters, quantized and pruned down from the original, running on a Kirin chip.

Trying it out, photographing the cat itself is really hard. I shot three rounds; the first two rounds the cat was moving, and the reconstructed ears came out blurry. The third round I got it while it was asleep. Generation took a few minutes, and it came out viewable by rotating, but don't expect fur texture.

Two pitfalls. One, the edge-side model is very picky about input — if the lighting is dim or the background is cluttered, the resulting model comes out floaty. Two, it only recognizes HarmonyOS; my own Android phone can't install it.

As for that 50FPS, I could only see it in the WAIC launch video. On site they used physical input to drive the frames. Frame rate — below 30 the human eye feels it stutter, 50 is a usable level. Inference cost reduced by 70% versus same-scale GPU solutions — I believe half of that number, since I haven't run it myself.

Conclusion

Depends. Suited for people who want to quickly make 3D models of physical objects and happen to use HarmonyOS, or teams doing embodied intelligence that need to run real-time simulation on domestic compute. Not suited for a student like me who only has an ordinary graphics card and wants to build a setup to play with.