01What Kimi K3 Is and How It Works

Kimi K3 is a new open source AI model from China that was released about 24 hours ago. It has 2.8 trillion parameters and a 1 million token context window. This means it can handle very long conversations or large documents at once.

Kimi K3 is designed to handle reasoning, coding, and vision tasks. It uses something called vision in the loop, where it can see its own work and improve it automatically. For example, if it creates a game, it can look at the screen, find issues, and fix them without you asking.

02Main Features of Kimi K3

Kimi K3 comes with several powerful features. First, it has a huge context window of 1 million tokens, which is bigger than Claude Fable 5. Second, it is open source, so anyone can download and modify it.

Third, it uses two new architectures called Kimi Delta attention and attention residuals, which help information flow better inside the model.

Other features include:

  • A 2.5x improvement in scaling efficiency compared to K2.
  • Self-improvement through a 15-hour nonstop iteration loop.
  • Support for tool use through Hermes Agent, including Blender MCP for 3D modeling.
  • Availability through Kimi.com, Kimi Work, Kimi Code, and Kimi API.

03How to Set Up Kimi K3

Getting started with Kimi K3 is straightforward. Follow these steps:

  1. Go to Kimi.com or use the Kimi API to access the model directly.
  2. If you want to use it with agents, download and install Hermes Agent from the GitHub repository.
  3. Create a new agent profile inside Hermes Agent and plug in the Kimi K3 API key.
  4. Set up an Obsidian memory system to store your preferences and past work.
  5. Connect any MCP tools you need, like Blender MCP for 3D modeling or a trending tool for social media.
  6. Test the setup by running a simple task, like generating a 3D game or a website.

04How to Use Kimi K3 with Examples

In this episode, Julian shows several real examples of what Kimi K3 can do. He created an open world game similar to Skyrim using a single prompt. The game has beautiful graphics, smooth controls, and a nice ambiance.

He also made a Minecraft style game, a Dragon Realm game, and a fully animated video using Reemotion and Hermes Agent.

For coding, you can use Kimi Code inside a workspace. All your projects are saved in one place. You can also use the CLI inside the chat and switch between quality, fast, and no reasoning modes.

For automation, you can create a custom workflow that analyzes trending data from Twitter, then automatically generates SEO and social media content in one click.

05Results and Comparisons from Tests

Julian tested Kimi K3 against Claude Fable 5 and GPT 5.6 using GoldieBench, a set of 50 different tasks. For a Minecraft style game, Kimi K3 produced a super nice result with water and smooth graphics. Fable 5 looked okay but had bugs and controls that were not as good.

GPT 5.6 came nowhere near, with much poorer quality.

For a Dragon Realm game, Kimi K3 again looked better than both competitors. Fable 5 lacked detail and ambiance, while GPT 5.6 felt like a whole generation behind. On benchmarks, Kimi K3 is in the top three for almost everything, and sometimes it outperforms GPT 5.6 and Fable 5.

Julian notes that Opus 4.8 is far behind Kimi K3 on all benchmarks.

06Tips, Limits and Mistakes to Avoid

When using Kimi K3, keep these points in mind. First, it is still new, so some tasks may not work perfectly yet. Julian says Kimi K3 is the worst it will ever be because it improves rapidly.

Second, if you use the API, it is much cheaper than Claude or GPT, but check the pricing for your specific use case.

Also avoid relying on Kimi K3 for everything. Some benchmarks still show Fable 5 outperforming it. Use Kimi K3 as one tool among many.

Finally, if you want to use it with Blender MCP or other MCPs, make sure your agent harness is set up correctly. You can also combine it with Claude Code to get the best of both worlds.

07Who Kimi K3 Is For and Why It Matters

Kimi K3 is for anyone who wants a powerful, cheap, and open source AI model. Developers can use it to create games, websites, or automation tools. Business owners can use it to generate content or analyze trends.

It is also great for people who want to experiment with the latest AI without paying high prices.

This model matters because it shows that open source models from China are catching up to and even beating frontier models from the US. It creates pressure on companies like OpenAI and Anthropic to improve faster and lower prices. For users, this means better AI tools at lower costs.

08How Kimi K3 Compares to Previous Models

Kimi K3 is a huge leap over its predecessor, Kimi K2.7, which came out only a month ago. The context window went from 262K tokens to 1 million tokens. That is a quadruple increase.

On benchmarks, there is a clear generational gap between K2.7 and K3.

Kimi K3 also improves on scaling efficiency by 2.5 times thanks to new attention mechanisms. It is much better at internal knowledge work and can self-improve for 15 hours nonstop. Julian says if you are still using Opus 4.8, you should switch to Kimi K3 immediately because it is far ahead on every test.

09Best Ways to Use Kimi K3 with Other Tools

You can get the most out of Kimi K3 by plugging it into Hermes Agent. Julian shows how it works with an agent operating system that runs side by side with Claude and GPT. You can use it with the Blender MCP to control 3D modeling software, or with a trending tool to automate content creation.

Another great option is to combine Kimi K3 with the Obsidian memory system. This lets the model remember your projects, goals, and voice across sessions. You can also use it with Claude Code's agent harness for additional flexibility.

Finally, you can create custom workflows that run on a 24-hour schedule using cron jobs with Hermes Agent.