This Open-Source AI Taking Over 3D & Blender — Kimi K3

Summary
Stefan Vaskevich (creator of Stefan 3D AI) evaluates Moonshot AI’s open-weights model Kimi K3 across a 10-test benchmark suite for 3D generation, animation, physics simulation, and game development. He compares Kimi K3 directly against Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol / Codex using command-line agent harnesses connected to Blender MCP, Godot MCP, and Fal.ai. Across isolated 3D asset generation and two playable game prototypes (a Mario Kart clone and a SimCity clone), Stefan demonstrates that Kimi K3 closely rivals or outperforms proprietary frontier models at a fraction of the inference cost.

What is shown

• [00:00–01:02] Arena benchmarks and context: Presenter reviews leaderboard placements, including Kimi K3 reaching #1 on the Frontend Code Arena (Elo 1,679) and DesignArena 3D Design (Elo 1,450), Moonshot’s subscription capacity pause notice, and Artificial Analysis pricing charts.
• [01:03–01:52] Testing environment setup: Stefan configures three CLI agent sessions running Claude Fable 5 (max thinking), Kimi K3 (max thinking), and GPT-5.6 Sol / Codex (ultra reasoning) connected via Blender MCP, Godot MCP, and Fal.ai MCP.
• [01:53–04:27] Test 1 (Blender Donut): Each model proceduralizes and renders a frosted sprinkle donut from a single 2D reference image; Claude and Kimi produce clean models while Codex generates odd, unnatural droplet geometry.
• [04:28–05:43] Test 2 (Diamond Shader in Godot): The models generate custom dispersion/refraction shaders on an imported diamond mesh; GPT-5.6 Sol produces the most visually dynamic refraction, while Kimi K3 creates a colorful, rainbow-like faceted effect.
• [05:44–07:12] Test 3 (Sci-Fi Hologram in Godot): Generating a hologram effect and animated pedestal for a GLB character; Claude and Kimi succeed with scanning lines and podium cones, while Codex’s floating halo/particle construct fails to align properly.
• [07:13–08:46] Test 4 (Procedural Three.js Clamshell Console): Models write procedural interactive 3D code using img2threejs for an opening/closing handheld console; Claude and Kimi generate functional, clicking, hinged devices, while GPT-5.6 Sol creates an incorrect non-hinged chassis.
• [08:47–10:41] Test 5 (Physics Simulation in Blender): Shattering a pane of glass with a rock to have fallen shards spell out “KIMI”; Claude and Kimi achieve readable lettering, whereas GPT-5.6 Sol throws shards backwards unnaturally.
• [10:42–12:13] Test 6 (Animation Retargeting in Blender): Retargeting a Mixamo dance animation onto an AccuRIG humanoid character; all three models succeed in a single prompt with accurate bone mapping.
• [12:14–14:00] Test 7 (Dragon Rigging & Flight Cycle): Rigging and animating wing-flapping on a solid dragon mesh; Claude suffers weight-painting tearing, GPT-5.6 Sol produces distorted geometry, while Kimi K3 produces the cleanest rig and flapping loop.
• [14:01–15:06] Test 8 (Toon Bonfire in Godot): Procedural stylized flame shader with smoke and ember particles; Claude produces stylized flame sheets, Kimi creates faceted flame shards, and Codex builds an elaborate voxel-style camp scene.
• [15:07–19:58] Test 9 (Mario Kart Clone in Godot): End-to-end game generation with beach and ice tracks calling Fal.ai for 3D kart and prop generation; Claude Kart has track boundary bugs, Codex Kart 64 lacks steering controls and textures, while Kimi Kart produces fully drivable drifting, AI racers, and functional race tracks.
• [19:59–23:25] Test 10 (SimCity Clone in Godot): Overnight 4–5 hour autonomous generation run by Kimi K3 building a SimCity 2000-style simulation; Stefan plays the resulting game, laying roads with spawning traffic, managing water/power plants, and zoning areas that autonomously grow into dense 3D skyscrapers.

Claims & numbers

• The presenter notes Claude Fable 5 was temporarily restricted a few weeks prior due to cyber safety capabilities (screen shows article dated July 6, 2026).
• The presenter states Kimi K3 took the #1 spot on the Frontend Code Arena leaderboard with an Elo rating of 1,679, beating Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618).
• The presenter highlights Artificial Analysis data showing Kimi K3 costs $0.95 per Intelligence Index task versus $3.75 for Claude Fable 5 and $1.80 for GPT-5.6 Sol, making it nearly four times cheaper than Fable 5.
• The presenter states Kimi K3 scored 1,450 on the DesignArena 3D Design benchmark, ranking #1 ahead of Claude Fable 5 (1,387) and GLM-5.2 (1,383).
• Claude Fable 5 took 1 hour and 10 minutes to compile the Kart game, while Kimi K3 required approximately 4 to 5 hours of compute to build the SimCity clone.

Notable quotes

• [00:00] “Who would have guessed that open source will take over that soon?”
• [14:48] “I like, again, like this guy is always try to add something I don’t ask.”
• [22:53] “How cool is that? I’m absolutely amazed.”

Assessment
This is a hands-on technical review and independent benchmarking video. All demonstrations show real CLI terminal logs, Blender 3D viewports, and Godot engine runtime executions, with genuine technical glitches, asset loading quirks, and simulation bugs transparently displayed and analyzed.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.