This is the English reading version of the Dutch news item Kimi K3.1 Pro maakt sterke claims voor VibeCad resultaten.
Moonshot AI has released Kimi K3.1 Pro, a model of 2.8 trillion parameters that according to the company itself is strong at 3D work. The announcement says it makes parametric models, from mechanical parts to furniture, and that it writes and runs scripts for Blender and CAD programmes.
The model uses only part of its computing power per question: of the 2.8 trillion parameters, 104 billion are active. It reads text, images and video, has a context window of one million tokens, and there is a faster variant. You use it through Moonshot’s own apps or through the API.
On BenchCAD, the benchmark that has a model write CAD code from several views of an object, Moonshot reports 92.7% with tools. That figure was measured by the company itself, on a random selection of 1,000 of the 17,900 tasks and in the environment Claude Code uses; BenchCAD’s own leaderboard also lists that line as self-reported. It puts it next to Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%), figures that, like this one, come from those makers’ own reporting. What stands out is that the model writes and runs those scripts without an MCP connection as well, so without the layer that CAD programmes today use to connect to a model. Its own app also has 3D Make: view and adjust a model in the browser, and download it as a print file.
The files with which you can run the model yourself are promised for 4 November, and the announcement states no licence yet. That is where the promise pinches: 2.8 trillion parameters with 104 billion active is not a model you run alongside your CAD programme on a workstation.
My verdict: the claims are impressive, we will see later whether they hold up.
Which CAD programmes already have an MCP connection is in the overview on this site. Keep an eye on this site for updates and tests.
The announcement is on Moonshot AI’s tech blog, the figure with its caveat on BenchCAD’s leaderboard.