Kimi K3.1 vs Qwen 4: Build a 3D game of someone riding a cow using Three.js.

This is the kind of coding prompt that reveals more than whether a model can generate a working 3D scene.

The funny part? Kimi K3.1 seems to have corrected the prompt itself.

Instead of blindly following the request, it appears to have reinterpreted what the user wanted. That might be a clever recovery from an unusual prompt, or it might be the model quietly deciding it knows better than the user.

And that distinction matters when you’re building with AI.

A good coding model should be able to handle weird ideas, translate them into working implementations, and recover when its initial interpretation doesn’t make sense. But it should also preserve the user’s actual intent rather than silently replacing it with a more conventional task.

Three.js makes this a fun test because the model has to do more than write ordinary application code. It needs to create a 3D scene, position and animate objects, handle camera movement, and potentially implement controls and collision logic.

Making a character ride a cow is ridiculous enough to be interesting, but technically specific enough to expose differences in implementation.

The real comparison is whether Kimi K3.1 and Qwen 4 can turn the same prompt into a playable 3D experience, and how much correction each needs along the way.

Sometimes the most revealing coding benchmark is a prompt nobody would put in a formal benchmark. https://x.com/TimJayas/status/2113277551980482975/video/1