这篇文章还没有中文版,下面显示的是英文原文。
Kimi K3 Explained: Can a 2.8-Trillion-Parameter AI Really Build a 3D Game?
Kimi K3's game demos, vision-in-the-loop coding, benchmarks and its mixture-of-experts design, with Kimi's own claims kept apart from what is verified.
Kimi K3's launch demos show an open world, a handheld game and other playable scenes, all generated by the model. The visuals are exciting. The real question is whether the model can write code, look at the result and improve it.
This is an independent look at Kimi's published material. I separate what Kimi claims from what anyone can check. I later ran my own test on the newer model: Kimi K3.1 Pro built a 3D game from one prompt.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
Website creation
You describe a page, show an image or give a visual reference. Kimi builds a working design and lets you change it by chatting. That speeds up early prototyping.
A generated website still needs a human check before it goes live: accurate text, responsive layout, accessibility, performance, privacy and security.
Games and "vision in the loop"
Kimi uses game development as its example of long, connected work. The model has to create terrain, lighting, movement, controls and a working camera, then run the project and look at it.
Kimi calls the method vision in the loop:
- Write the code.
- Capture a live screenshot.
- Compare it with the goal.
- Fix what looks wrong.
This is useful for prototyping. But a polished recording does not prove the whole project is stable. Test the controls, performance, saved states and code quality yourself. The real test is whether another developer can open the project, reproduce the result and keep working on it.
Benchmarks
Kimi publishes charts that compare K3 with other models on software engineering, terminal tasks, programming problems and internal evaluations.
Keep three things in mind:
- A benchmark is not a production project.
- Scores change with the prompt, tools, time limits, reasoning settings and grading method.
- Kimi says these results use maximum thinking effort. They are company-published numbers under specific conditions.
My preferred test is practical: give the model a real repository, clear requirements, version control and automated tests. Then check that it solves the task without breaking what already works.
Knowledge work
Kimi also presents K3 for research, reports, presentations and tasks that pull from many documents. A large context window helps the model keep more material together. It does not guarantee correct conclusions. Check important numbers, citations and dates against the original source.
The architecture, in plain words
| Term | What Kimi says it does |
|---|---|
| Kimi Delta Attention (KDA) and Attention Residuals | Extra connections that help information move across many layers, not only along one standard residual stream |
| Mixture of experts | Many specialist components exist, but only a small group runs for each token |
| Expert routing | Scores many experts, then picks a few. Kimi reports 16 experts active per token from a much larger pool |
| Quantile balancing | Shares the work more evenly, so a few experts are not overloaded while others sit idle |
The goal is high capacity without using every parameter at every step. Quantile balancing is a training and design idea, not a setting you control, but it can affect speed, capacity and quality.
Diagrams explain a design. They do not prove a better experience. What matters is whether the system follows instructions, uses tools safely, keeps context and finishes long tasks reliably.
Limits Kimi admits
Kimi itself lists limitations, including sensitivity to the model's thinking history and the chance of unexpected decisions. For important tasks, use checkpoints, backups, tests and human approval before anything destructive.
Verdict
Kimi K3 looks ambitious, most of all for visual coding and long agent tasks. Treat the demos and charts as Kimi's explanation. Independent testing is still needed for real-world performance, cost and reliability. AI can speed up the first draft. You stay responsible for accuracy, security and the final decision.
Links
- Full video: https://youtu.be/FH9ulnIys8M
- Kimi K3 technical blog: https://www.kimi.com/blog/kimi-k3
- Kimi website creation: https://www.kimi.com/features/websites
- My hands-on follow-up: Kimi K3.1 Pro game test
Independent review. Not sponsored by or affiliated with Kimi or Moonshot AI.


