Matthew Berman tests Kimi K3.1 using a Rubik’s Cube simulation, a phone rendering, a knot in Blender and a browser game. He praises the cube but finds inaccuracies in the phone and knot outputs and visual issues in the game.
He examines a website redesign and a generated video, then demonstrates that a benchmark graphic initially included invented placeholder figures. He supplies source numbers and describes needing additional prompting to get the requested image.
Berman discusses the model’s published specifications, benchmark comparisonsA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. and pricing. He regards it as a strong open-weights model, while saying his own testsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. do not match the strongest proprietary models.
Watch on YouTube
