01What Is Kimi K3 and How Does It Work?
Kimi K3 is a new AI model from the Chinese company Moonshot. It was released just about an hour ago and is already showing impressive results in coding and 3D game building. This model has 2.5 trillion parameters and a 1 million token context window, which is four times bigger than the previous Kimi K2.6 model.
Kimi K3 works like other large language models but is designed for coding tasks. You can use it in a command line interface (CLI) by getting a coding plan. It is also built into an agent operating system, which gives it access to memory and workflows.
This makes it useful for building complex projects like games and websites.
02Main Features and What Is New in Kimi K3
Kimi K3 has several big upgrades from earlier versions. It has 2.5 trillion parameters, which is a lot more than most models. Its context window is 1 million tokens, so it can handle very long conversations or large code files.
The previous model, Kimi K2.6, only had 256K tokens.
The model is also very fast and cheap. Julian Goldie says the coding plans are way cheaper than other frontier models. This means you get high performance without paying high prices.
Another new thing is that Kimi K3 is expected to be open source, like Kimi K2.7 was. This makes it accessible to more developers.
In benchmarks, Kimi K3 beats Opus 4.8 and is close to Fable 5 and GPT 5.6 Soul. On the Terminal Bench 2.1, it even surpassed Fable 5 and Opus 4.8. The model also jumped from number 18 to number 1 in the Front End Code Arena, beating Claude Fable 5.
03How to Set Up Kimi K3 for Coding
To use Kimi K3, you need to get a coding plan. This is cheaper than plans for other frontier models. Once you have the plan, you can plug it into a CLI and start building.
Here are the steps:
- Go to the Moonshot website and sign up for a Kimi K3 coding plan.
- Open your command line interface (CLI) on your computer.
- Connect Kimi K3 to the CLI using the provided API key.
- Optional: Plug Kimi K3 into an agent operating system to use it with memory and workflows.
- Start giving it coding tasks, like building a game or a website.
04Using Kimi K3: Game Building and Website Design Examples
In the episode, Julian tested Kimi K3 on several tasks. One example was building a 3D racing game called Neon Racer. The new version made with K3 looked much better than the old version made with Kimi K2.7.
The difference in quality was huge.
Another test was website design. Straight out of the box, Kimi K3 gave generic AI fluff – websites that looked okay but not special. However, when Julian used a web design skill with the model, the results were 10 times better.
The full blog generated with K3 looked really nice.
Julian also used Kimi K3 with Hermes Agent. He created a new profile and plugged K3 into the agent. It worked smoothly and replied fast.
This setup lets you use the model for agentic tasks, which Claude does not allow on a subscription plan.
For building tools, Kimi K3 was used to create a racing game and a blog. Both looked impressive. The model also performed well on the BridgeBench Horror House game test, beating Fable 5.
05Benchmarks and Comparisons: Kimi K3 vs GPT 5.6 and Fable 5
Kimi K3 was tested against other top models. On the Frontier SWE benchmark, it scored 81.2, compared to Fable 5's 86.6. Fable 5 was still ahead, but K3 beat Opus 4.8 and was close to GPT 5.6 Soul.
On Terminal Bench 2.1, Kimi K3 was right up there with GPT 5.6 Soul and surpassed Fable 5 and Opus 4.8. These numbers came from Leo's tweets. The model also ranked number 1 in the Front End Code Arena, surpassing Claude Fable 5.
This was a 17-place jump from Kimi K2.6, which was at number 18.
On the BridgeBench Horror House game test, Kimi K3 beat Fable 5. Julian also noted that the model is a huge step up from Kimi K2.7. The old model could barely create a decent racing game, but K3 made one that looked insane in comparison.
06Tips, Limits, and Mistakes to Avoid with Kimi K3
Kimi K3 is powerful, but it has limits. For website design, straight out of the box, it gives generic AI fluff. To get better results, you need to plug in your own web design skills first.
This means providing detailed instructions or using a skill set.
Another limit is that the model is very new. Julian said he had not had time to test everything. So you may find some tasks where it does not perform as well.
Also, while it beats Opus 4.8, it still falls behind Fable 5 on some benchmarks like Frontier SWE.
A tip is to use Kimi K3 with an agent like Hermes Agent. This gives you access to workflows and memory, making the model more effective. Also, the model is cheap, so you can run many tests without spending a lot.
Avoid expecting perfect websites without extra guidance. For 3D games and coding, it works great.
07Who Should Use Kimi K3 and Why It Matters
Kimi K3 is for developers and anyone who builds with AI. If you need a cheap but powerful coding model, this is a good choice. It is especially useful for building 3D games and apps that need a large context window.
Julian says Chinese open source models like Kimi K3 are getting better and better. They release new models each month. This makes frontier-level AI available at low cost.
If you are in a community like AI Profitable Boarding, you can get the full system with K3 plugged into an agent OS.
Why it matters: Kimi K3 shows that open source models can compete with the best paid models. It beats Opus 4.8 and comes close to GPT 5.6 and Fable 5. This is a big step for AI accessibility.
It also proves that Chinese AI companies like Moonshot are serious competitors.
