01What is Kimi K3 and How It Works

Kimi K3 is an open-source AI model from the Chinese lab Moonshot AI. It can create games, videos, and applications from simple text prompts. In this episode, Julian Goldie tests it against Claude Fable 5 and GPT 5.6 across many tasks.

The model works by generating complete outputs, like full game levels, in one go. It uses an API and can be plugged into tools like Hermes or Claude Code. Because it is open source, you can run it with a coding plan and avoid expensive token costs.

Julian emphasizes that Kimi K3 feels smooth and responsive in every test. He believes open-source models like this are now at the frontier of AI development.

02Main Features and What’s New in Kimi K3

  • Open source: You can use it freely, unlike GPT 5.6 and Fable 5.
  • Cheap: Input tokens cost $3 per million, much lower than GPT 5.6.
  • Fast progress: Kimi releases new models almost monthly, improving 2-3 times per generation.
  • Multi-modal: Can create 3D games, videos, and text responses.
  • Works with agents: You can plug it into Claude Code, Hermes, or an agent operating system.

Julian notes that Kimi K3 is the first model to complete some tests in a unique way, like creating a dragon style never seen before. It also handles 50 tasks in under 24 hours after release.

03How to Set Up Kimi K3

  1. Get a coding plan that includes API access to Kimi K3.
  2. Choose a tool like Hermes, Claude Code, or an agent operating system.
  3. Connect Kimi K3’s API key to that tool.
  4. For agent use, import the model into your existing agent setup.
  5. Start generating text, games, or videos by writing prompts.
  6. If you need to compare models, open a group chat with multiple agents.

Julian says you can also use Kimi K3 inside Claude Code by switching between models. He recommends having an agent operating system to manage all your AI agents in one place.

04Using Kimi K3 for Games and Videos

Julian tested Kimi K3 on several games: Skyrim-style open world, Dragon Realm, racing, Neon City, a crit maze, a black hole simulation, and fluid in a box. In most cases, Kimi K3 produced the most detailed and smooth graphics.

For example, in Dragon Realm, Kimi K3 created a dragon unlike any other AI’s output. The snow, sky, and controls felt natural. In the racing game, the colors and vibe stood out, with smooth gameplay.

Even in Neon City, the city blocks and lighting were impressive.

For videos, Julian used Kimi K3 with a skill called ReMotion to create a fully automated video with camera angles. The video even moved as a background when he moved his mouse—a unique effect he had never seen before.

05Results and Comparisons from the Episode

Julian compared the three models side by side. Kimi K3 won on graphics in every game test. GPT 5.6 often made gameplay more interesting, like adding enemies, but had problems like reversed buttons in the Skyrim test.

Fable 5 felt basic or buggy in most cases.

Benchmarks from Leo show: Terminal Bench 2.1, Kimi K3 scored 88.3, GPT 5.6 scored 88.8, and Fable 5 scored 84.6. Julian notes that Kimi K3’s previous version, K2.7, was far behind but improved hugely in one month.

Price comparison: Kimi K3 costs $3 per million input tokens, while K2.7 cost $0.72 per million. The quality jump is 2-3 times better, according to Julian.

06Tips, Limits, and Mistakes to Avoid

  • Don’t trust benchmarks blindly: Julian tests himself because benchmarks can be too optimistic or pessimistic.
  • Watch token limits: GPT 5.6 and Fable 5 ran out of tokens on subscriptions, but Kimi K3 did not.
  • Be ready for failures: Some Kimi K3 tests crashed, like the fluid in a box, but regenerating works.
  • Use the right model for the job: For big projects where mistakes matter, Julian still prefers Claude for reliability.

Avoid using Kimi K3 for complex tasks if you need absolute stability; it may fail occasionally. But for creative and cost-sensitive work, it’s a great choice.

07Who Should Use Kimi K3 and Why It Matters

Kimi K3 is for creators, developers, and businesses that want high-quality outputs without high costs. It’s especially useful if you build games, websites, or videos. Because it’s open source, you can integrate it into your own tools.

Julian emphasizes that open-source models are evolving faster than commercial ones. He expects a new Kimi or GLM model every month. This means more innovation and lower prices for everyone.

If you want to maximize your results, combine Kimi K3 with other models like Fable 5 or GPT 5.6 in a multi-agent setup. This way, you get the best of graphics, gameplay, and reliability.