Kimi K3.1 Pro is Moonshot AI’s hosted model, which the company says can create and refine 3D assets by writing and running Blender or CAD scripts. A separate feature, Kimi 3D Make, lets users preview and edit models online, download them for 3D printing, and export videos or Live Photos. The distinction matters: one workflow uses scripts to build or refine assets, while the other provides an online space to work with models.
What Kimi 3D Make offers
Moonshot describes 3D Make as a feature for previewing and editing models online, downloading them for 3D printing, and exporting videos or Live Photos. The company said it was live on Kimi.ai and the Kimi app.
How K3.1 Pro creates and refines 3D assets
Moonshot says K3.1 Pro can use a text brief, sketch, or reference image to infer an object’s shape, proportions, and spatial relationships. It can then write and run Blender or CAD scripts to create or refine assets, adjust geometry and materials, and arrange a scene.
That script-based workflow is distinct from 3D Make’s online preview and editing tools. Moonshot says its Blender and CAD scripting does not require MCP connectors.
What Moonshot’s BenchCAD scores measure
Moonshot reports Kimi K3.1 Pro scores of 92.7% with tools and 72.8% without tools on BenchCAD. The benchmark tests reconstruction of 3D objects from multiple rendered views by generating CAD code. Moonshot says the open-model results used a randomly sampled subset of 1,000 cases from a 17,900-case dataset. Its with-tools runs used the Claude Code harness; its without-tools results came from single-shot generation.
| Test setup | Kimi K3.1 Pro score | What the setup means |
| With tools | 92.7% | Reconstruction through CAD code generation using the Claude Code harness |
| Without tools | 72.8% | Single-shot generation using the official prompts and scoring pipeline |
The scores describe performance on this particular reconstruction benchmark under two different setups. They do not measure every kind of 3D task.
What creator demonstrations show
Viktor Oddy’s screen-recorded demonstration shows Kimi K3.1 Pro generating interactive 3D websites from prompts and Pinterest references. It includes a second example involving a real-estate space and a side-by-side website-generation segment with GPT-6 Astra. The video is a creator-run demonstration, rather than a standardized benchmark.
A separate, single sequential AgenC Core run compared Kimi K3.1 Pro and DeepSeek V4.1 Flash on the same detailed V8-engine task in Three.js, a JavaScript library for 3D graphics in the browser. Reddit user pzarevich reported that both produced working interactive simulations at maximum effort. Kimi finished in about 31 minutes after using 3.5 million total tokens; the harness estimated its cost at $2.16. DeepSeek was stopped after about 48 minutes while still refining, after browser interaction checks had passed. Its reported total was 67.2 million tokens, including 66.7 million cached input tokens, and the harness estimated its cost at $0.86. Those figures describe one run and its harness estimates, not a standardized cost or quality comparison.
NeoTeo’s earlier Kimi K3.1 Pro coding-test coverage examined software-engineering tasks, a separate question from 3D reconstruction.
Highspeed accelerates model output, not every step
Moonshot says Kimi K3.1 Pro Highspeed produces model output about two to three times faster. Kimi Code’s documentation says Highspeed uses about twice the quota, while tool calls and script execution are not accelerated. A faster response from the model therefore does not mean an entire 3D task takes two to three times less time.
Kimi Code lists regular K3.1 Pro access from the Plus plan and Highspeed access from Pro or above. Its 1,048,576-token context window is listed for Pro or above.
Hosted access and the planned weights release
On October 23, 2026, Moonshot announced that K3.1 Pro and Highspeed were rolling out through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The company separately planned to release the full model weights by November 4, 2026, distinguishing that planned release from access through its hosted services.
