Kimi K3.1 Pro: an open-weight model pitched at Blender, CAD and Three.js scenes
Moonshot's new 2.8T-parameter MoE (104B active, 1M context) is marketed for writing and running Blender and CAD scripts and for animated Three.js scenes. Weights are promised by Nov 4.
Moonshot launched Kimi K3.1 Pro on Oct 23. According to the launch post and blog it's a 2.8T-parameter MoE with 104B active parameters and a 1M-token context window, available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with a 2–3× faster 'Highspeed' variant. Open weights are promised by Nov 4. The launch post passed 18k likes within hours, and 3D creators started posting tests the same afternoon.
The video angle is 3D. The blog says K3.1 Pro writes and runs Blender and CAD scripts without MCP connectors and builds Three.js scenes with stronger animation. It reports BenchCAD at 92.7% with tools, against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra, and Terminal-Bench 4.0 at 65.7%, against 66.4% and 58.2%. Kimi.ai also has a '3D Make' feature for previewing and editing 3D models and exporting them as video or Live Photos. Community tests so far: @crystalsssup posted a Changi Airport scene modeled, assembled and rendered in Blender, and @animiotics put the same 40-second science animation through K3.1 Pro in Blender and Claude Opus 5.5 side by side.
Why it's on ReelStack: agent-driven Blender blockouts, procedural product shots and JS motion graphics (Three.js, HyperFrames, Remotion) all depend on how well a model writes scene code. If K3.1 Pro gets close to Opus at a lower token price, these pipelines get cheaper to iterate on. @k1rallik's own Roadster comparison puts it at about $2 against $15 for Opus 5.5, but those are his numbers, not a benchmark.
Our take: the benchmarks are Moonshot's own. Some 3D artists are already asking whether 'good at 3D' means production assets or browser-demo Three.js. Run it against your current agent on one real shot before switching.
More from the Lab
- Fable 5.2 turns 2D floor plans into walkable 3D tours: the archviz workflow taking off on X
- anymove: creature mocap from one two-view AI video
- Code2Video-AA ranks 27 models on writing videos in code: Claude Opus 5.5 leads, GPT-6.1 Sol is the value pick
- Block the shot in Blender, render it in LTX: the Layout-to-Render LoRA
Topics: Open-weight video & audio models