← K3 overview

First Test of Kimi K3 — Building 3D Games with Swarm Agents

2026-07-17 — Michael Leung  ·  #kimi#k33D gamesagentic coding

This is the short write-up of my first hands-on test with Kimi K3, the new 2.8-trillion-parameter open-source flagship from Kimi. The verdict up front: according to my research on the internet, this model is very good at creating 3D games — so I put that claim to the test and used K3 swarm agents to build one. The result was pure agent code — a playable 3D web game after just 2 hours of coding with the swarm agent.

Why 3D games?

When a new model drops, I like to test it on something harder than "write me a sort function". Browser-based 3D games are a great benchmark for agentic coding because one prompt touches a lot of surface area at once: scene setup, rendering loops, camera controls, physics, collision handling, game state, UI overlays such as a difficulty selector, and the visual polish that makes it feel like a game rather than a tech demo.

The official K3 documentation backs this up as a target use case: K3 is built for long-horizon coding — sustaining long-running engineering tasks with minimal human supervision, working with large codebases, and coordinating terminal tools. It also combines software engineering with visual reasoning, using screenshots and visual feedback to improve workflows in game development and frontend engineering [official quickstart].

The setup: K3 swarm agents, pure agent code

Instead of hand-writing scaffolding and asking the model to fill in gaps, I ran the build through K3 swarm agents and let the agents own the whole loop — planning the structure, writing the code, running it, looking at the result, and iterating. My role was closer to a product owner than a programmer: describe the game, set acceptance criteria, review the output.

• Goal: a standalone 3D web game, playable in the browser.
• Constraint: pure agent code — no hand-written starter template.
• Features requested: a difficulty-level selector, real-time 3D rendering, and playable game logic.
• Time invested: about 2 hours of coding with the swarm agent, from first prompt to playable build.
• Process: prompt → agent build → visual check → refinement loop.

Two hours, start to playable

I want to be honest about the timeline, because it is the part that surprised me most. I created the game after roughly 2 hours of coding with the swarm agent — and most of my side of those two hours was writing the spec, watching the agents work, and playing the intermediate builds to give feedback. The agents did the typing.

In that time the swarm produced the full project: scene and rendering setup, controls, game logic, the difficulty selector, and the UI polish around it. No starter template, no copy-pasted tutorial code — every line came from the agents. Two hours from an empty folder to a playable 3D game is a workflow I simply did not have before K3.

What impressed me

The standout was not any single snippet — it was the stamina. K3 held the context of the whole project across a long session, kept earlier decisions consistent, and used visual feedback (screenshots of the running game) to fix issues that are painful to describe in words, like camera angles and object placement. That is exactly the "long-horizon" behaviour the release notes promise, and it is what makes 3D work practical from a chat window.

My honest opinion: K3 can beat Fable 5 at coding

Here is my personal take, and I say this as someone who has been a programmer for over 20 years: I believe Kimi K3 can beat Fable 5 (Claude Fable 5) in terms of coding.

Fable 5 is currently the model everyone measures coding agents against, and fair enough — it is excellent. But benchmarks are one thing and shipping is another. In this test, K3 sustained a real, multi-step build for 2 hours, kept the whole project coherent, used visual feedback to fix its own work, and delivered a playable game with zero hand-written code from me. That is the exact job description of a top coding agent — and K3 did it as an open-source model that I can also call flat-rate through the API, with a 1M-token context and automatic caching.

Is that a scientific benchmark? No — it is one developer's hands-on test. But if an open model gives me the same practical result I would expect from the most expensive proprietary coding model, then for my day-to-day work the crown is up for grabs. I will keep testing and update this page as I push K3 harder.

Try it yourself

If you want to run the same experiment, the Kimi K3 quickstart on this site has everything you need: an API key, the OpenAI-compatible endpoint, and working examples for streaming, vision input and tool calling. Give the agents a clear spec and let them cook.