Summary
Matthew Berman puts the newly released Kimi K3.1, a roughly three-trillion-parameter open-weights model from Moonshot AI, through a series of hands-on tests after subscribing and installing the Kimi desktop app.
The first test is a 3D Rubik’s Cube simulation, which open-weights models have historically failed. Kimi K3.1 produced a realistic, interactive cube with scrambling, keyboard controls, a timer, and an algorithmic solver, which Berman calls one of the best results he has seen. A WebGL render of the iPhone Duo looked polished but was inaccurate, took nearly 29 minutes, and did not improve much after a follow-up prompt asking the model to research the real specs.
Berman also compares Kimi with GPT 6.1 Soul on a motion graphic explaining how to tie a belay knot. GPT 6.1 Soul’s version was clear enough to follow, while Kimi’s was not, even after a second pass. Another test, a Fall Guys clone for the browser, took 48 minutes to produce a playable game called Bean Blitz. Along the way he notes Kimi’s frequent “eyebrow text” style, similar to Claude models, and mentions accusations of distillation. The takeaway: strong 3D and front-end ability for an open model, with gaps against the closed frontier on some tasks.
📺 Source: Matthew Berman · Published October 24, 2026
🏷️ Format: Review
