KIMI K3 is Ridiculous.

Summary
In this video, creator tef tests Moonshot AI’s Kimi K3 model inside the Kimi Code environment in VS Code. Prompting the model at maximum thinking effort, the creator benchmarks its ability to autonomously code three complex 3D browser games from scratch—clones of Minecraft, Fortnite, and Grand Theft Auto—comparing its capabilities and output to Anthropic’s Claude Fable 5.

What is shown

• [00:04] Arena leaderboard showing Kimi-K3 ranked #1 with an Arena score of 1,679, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618).
• [00:10] The Kimi Code interface in VS Code, setting the thinking effort setting to “Max”.
• [00:40] Crafting and submitting a prompt to generate a full 3D voxel game called “Blockfront” using Three.js/WebGL.
• [01:29] A promo card detailing Kimi K3 specifications (2.8 trillion parameters, 1M-token context).
• [02:02] Kimi K3’s chain-of-thought planning and multi-file code generation (world generation, player physics, rendering shaders, UI, audio).
• [02:59] Gameplay demonstration of “Blockfront”: procedural terrain, biome generation (snow, beach, plains), animated swaying grass, pig/cow mob models, dynamic day-night cycle, custom block textures, and hostile zombie mobs.
• [06:37] Submitting a prompt for a complete battle royale game modeled after Fortnite.
• [07:20] Gameplay demonstration of “Dropzone Royale”: battle bus drop, skydiving/gliding, chest looting, weapons (SMG, shotgun), building mechanics (placing walls, floors, ramps to “crank 90s”), a shrinking storm circle, and elimination screens.
• [10:10] Submitting a prompt to build an open-world sandbox city game called “Sunset City”.
• [10:55] Gameplay demonstration of “Sunset City”: 3D urban environment, pedestrian movement, gun mechanics, police aggression with sirens and wanted stars, and drivable vehicle physics.

Claims & numbers

• The presenter highlights that Kimi K3 is ranked #1 on the Arena leaderboard with a 1,679 score, above Claude Fable 5’s 1,631 score.
• A graphic shown states that Kimi K3 has 2.8 trillion parameters and a 1-million-token context window.
• The presenter claims they spent $50 ($40 subscription plus $10 in usage credits) to generate the Minecraft clone project.
• The presenter states that generating Minecraft (“Blockfront”) took approximately 2 hours of processing and self-testing time.
• The presenter states that generating the Fortnite clone took 2 hours.
• The presenter states that generating the GTA clone took about 1 hour.
• The presenter rates the generated Fortnite clone an 8 out of 10.

Notable quotes

• [01:30] “2.8 trillion parameters, bro. One million token context. Like, what?”
• [02:49] “So I just spent $50 to make this Minecraft.”
• [12:18] “Guys, we just made GTA 6 before GTA 6 with Kimi K3 AI.”

Assessment
This is an independent user review and hands-on coding demonstration showing long-horizon agentic game generation in real browser environments. While the generation wait times (1–2 hours each) are jump-cut for brevity, the resulting playable WebGL games, complete with functional mechanics and controls, are demonstrated directly on screen.

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