Breaking

Moonshot AI's Kimi K3 Builds a Game-Ready Fighter Jet in Blender From a Single Prompt

July 18, 2026 at 19:15 EDT

AI Agents
Media Generation
Foundation Models

Moonshot AI's newly released Kimi K3 model produced a PlayStation 3-era fighter jet 3D asset inside Blender from one natural-language prompt, completing the job in roughly 25 minutes for less than a dollar in compute. The demonstration, which circulated widely among developers, has been held up as a striking illustration of how quickly the barrier to entry for 3D asset creation is falling.

July 16, 2026 · Moonshot AI

Kimi K3 Modeled a Fighter Jet in Blender From One Prompt

Driving Blender directly through the Model Context Protocol, Moonshot AI's flagship open-weights model blocked out a PS3-era game asset — rendering, inspecting and iterating on its own geometry — in about 25 minutes for under a dollar in compute.

~25 min

To complete the fighter-jet blockout, start to finish

< $1

Total compute cost for the full modeling session

2.8T

Parameters — first open-weights model in the ~3T class

The Agentic Loop

What separates this from earlier text-to-3D: the model runs Blender's bpy API like a human operator, then checks its own work.

ACT

Run bpy code to build geometry

→

RENDER

Produce an image of the scene

→

OBSERVE

Inspect its own output

→

REVISE

Iterate until the shape holds

API Pricing — Output Costs 5× the Input

Cost per million tokens. Low rates are what keep an entire modeling session under a dollar.

Input

$15

Output

Each block = $3 / million tokens. Output is 5× the price of input.

The Win

Collapses raw-geometry production into a single prompt — a meaningful accelerant for prototyping, turning hours or days of blockout work into a lunch break.

The Caveat

What it delivers is a blockout, not a shippable asset. Clean topology, UV unwrapping, textures and deformation-ready edge flow still require human or tool-assisted work.

A shift from manual modeling toward natural-language prototyping.

The open question is no longer whether an agent can rough out a mesh — it can — but how much of the finishing pipeline it can eventually automate as full weights and MCP-style tooling mature.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.