Same three.js Prompt, DeepSeek V4.1 Beats Kimi K3.1

Ivan Fioravanti posted a comparison on X: the same three.js prompt was given to DeepSeek V4.1 1017 (hereafter DS41) and Kimi K3.1 Pro, asking them to generate a freely walkable procedural world in the browser. The result was stated plainly: DS41 wins.

The prompt was provided by @sunglassesface and the requirements were not low: endless, procedurally generated, every area random, surprises everywhere; players can walk freely, encounter entities and interact; the visuals should be cool, while the overall experience should be calm, relaxing, and genuinely fun. The original prompt also added that it wanted the cozy satisfaction of a supermarket simulator, but it must not be a supermarket game.

Each model delivered a world. The first video is Kimi's StillWild. The scene has green grass, trees, and small animals; a character in orange clothes stands on a path; in the distance there is a pink tree; the mood is quiet.

The second video is DeepSeek's MossLight. In the screenshot there are floating islands, green grass, a character wearing a green hat and an orange scarf with their back to the camera, accompanied by a black animal companion. The text on screen reads: an endless countryside, suited to slow wandering; no map to fill in, no time limit; every hill and every stranger is a new adventure.

Ivan's assessment of the Kimi version was very specific: there were technical issues, such as the frame rate dropping from time to time, and overall it was not that cool either. This feedback is more interesting than the conclusion of 'who is stronger.' With an open-ended prompt, what the model generates is a complete runnable scene, and frame rate, camera, and sense of interaction are all exposed in the video. Leaderboard scores are one thing; whether frames drop on camera is another.

The replies did not stop at who won. Someone asked Ivan which DeepSeek V4.1 recipe and weights he ultimately used. He answered that he was still trying several: DeepSeek V4.1 Flash 1017 EXL3 2.9 bpw was not bad, fast, and its coding ability was still intact; on Apple Silicon he used DwarfStar Q4, which was a bit slower but whose quality satisfied him. He also asked the other person whether they wanted the DGX or Apple recipe.

Another replier pointed out a commonly overlooked variable: Was each model run once, or was the best picked from multiple runs? With this kind of open-ended prompt, results vary a lot from run to run. A single comparison cannot be treated as an absolute ranking.

A netizen said in the replies that DS4.1 1017 was amazing, and it was still a Flash model. There is a counterintuitive point here: Flash is usually easily regarded as the 'fast but weak' version, but at least in this kind of coding-plus-realtime-rendering task, it did not seem weak.

So what is really worth looking at in this comparison lies elsewhere: under the same prompt, what kind of 'walkable world' the two models made, whether the frame rate is stable, and whether the visuals make people want to keep walking a couple more steps. And the weights, quantization, and hardware behind the models all enter the final experience.

Source: https://x.com/ivanfioravanti/status/2113890712097825112