Kimi 3D Make: AI builds 3D models, games and print files from a chat

What it's about

Moonshot AI, the Chinese company behind the Kimi assistant, unveiled 3D Make on October 30, 2026. You describe what you want to build in plain language, or upload a sketch or reference image. Kimi then produces 3D models, interactive 3D web pages, browser games or objects you can 3D-print. This works because Kimi doesn't just output text. It operates professional software such as Blender and CAD programs. The feature is live on kimi.ai and in the Kimi app. Advanced workflows with locally installed software run through Kimi Work (a desktop app for Windows and Apple Silicon Macs) or Kimi Code.

How Kimi builds the models

According to Moonshot's technical guide, Kimi takes one of two routes depending on the object:

• Parametric modeling: For things that can be described with dimensions and rules (furniture, machine parts, buildings, product prototypes). The Kimi K3.1 Pro model writes Python scripts that control Blender or CAD software. The upside: you can adjust sizes and proportions precisely later.
• Mesh-based modeling: For characters, animals, statues and other organic shapes. Here Kimi prepares prompts and reference material and hands them off to specialized 3D generators such as Tripo, Meshy and Hyper3D. (A mesh is a net of many small triangles that describes an object's surface.)

The feature is split into three areas:

You can view models online, change them with follow-up instructions and download them.

Games and larger scenes

Moonshot says Kimi can also work with game engines such as Unreal Engine, Unity and Godot: assembling assets, building environments and programming game logic. As an example, the report mentions a small spaceship game with an asteroid field, obstacles and scoring. To launch K3.1 Pro, the company also showed a celestial palace inspired by Chinese mythology in Unreal Engine 5.8. Its source models reportedly contain around 36 million triangles, and it uses Nanite (detailed geometry) and Lumen (lighting and reflections). This is a vendor demo, not an independently verified measurement.

The model behind it: Kimi K3.1 Pro

K3.1 Pro was released on October 24, 2026 as the first model in the new K3.1 series, along with K3.1 Pro Highspeed, a variant that is 2–3 times faster with the same capabilities. It is a post-trained version of Kimi K3 with far more compute spent on reinforcement learning (learning through reward). Moonshot calls it its most capable "3T-class" open-weight model to date (3 trillion parameters), with a 1-million-token context window. According to Geeky Gadgets (based on a video by Matthew Berman), it is a mixture-of-experts model that activates 104 billion parameters per task and processes text and images. The Kimi Code documentation says it also accepts video.

Benchmarks according to Moonshot:

In an internal hallucination test, the share of false individual claims fell from 23.4 to 8.3 percent. Moonshot stresses that the test is deliberately error-prone and not a measure of everyday use. Geeky Gadgets ranks the model ahead of Meta's Muse Wonderful Full and Google's Argon at 3D modeling, but behind GPT-6.1 Soul and Opus 5.5 overall. It says the model is weaker on science benchmarks and tricky 3D rendering (such as tying knots).

Cybersecurity and open weights

Moonshot says it removed cyber-related data from training for the open-weight version. A checkpoint without these measures solved 90.5 percent of tasks in the company's own xdev-bench-hard test (pass@3). The final version solved 81.0 percent. Moonshot says its capabilities still trail those of proprietary frontier models that have been available for at least three months. The weights are due by November 4, 2026. Whether you can run them on a home PC is doubtful at this size. Memory requirements and the license are still to come.

Access and cost

K3.1 Pro is available in Kimi.ai, the apps, Kimi Work (from version 3.1.0), Kimi Code and via API (kimi-k3.1-pro, kimi-k3.1-pro-highspeed). In Kimi Code, the standard model requires a Plus plan, and Highspeed (which uses roughly twice the quota) requires a Pro plan. The 1-million-token window is reserved for Pro members. The model runs only in thinking mode. 3D Make draws on the shared Kimi usage quota. Details are in your account settings.

Limits

Moonshot itself admits that some projects need several rounds of revision, and that character rigging, complex materials and physics simulation are still weak. Check dimensions and geometry before any technical use, and check for holes and weak spots before 3D printing. There is no independent 3D benchmark. According to Geeky Gadgets, web design results also showed typical AI artifacts.

Our take

3D Make shows a trend: assistants that operate specialist software instead of just answering. For beginners, prototyping and comparing ideas (furniture, shop fittings, game assets), that's appealing. Professionals will still need to check and optimize the output. In CAD reconstruction, Kimi trails Claude Opus 5.5 and GPT-6 Astra slightly by its own numbers. Almost all the figures come from Moonshot itself. The picture will only firm up with independent tests and the released weights.