Video tutorial · The Metaverse Guy
Kimi K2 Thinking: The Open-Source Model That Outperforms Sonnet 4.5 & GPT-5
I put Kimi K2 Thinking through real-world coding and reasoning tests – it matched or beat Claude Sonnet 4.5 and GPT-5 while being faster and cheaper.
- Published
- 10 November 2025
- Length
- 13 min
- Views
- 17,228 views
In this article, I test Kimi K2 Thinking, a new open-source model from Moonshot AI. It impressed me with its reasoning speed, logic, and cost-efficiency, outperforming Claude Sonnet 4.5 and GPT-5 in several scenarios. I used it to extend a kids storybook generator with a new voiceover control and also ran website-based tests for voxel art generation and a financial research report.
Introduction to Kimi K2 Thinking
Kimi K2 Thinking is the most impressive open-source LLM I’ve tested so far. It’s fast, logical, and actually beats Claude Sonnet 4.5 and GPT-5 in reasoning and cost-efficiency. In this video, I wanted to see how it performs in real project work, not just benchmarks.
Website Tests: Voxel Art and Stock Market Research
Before diving into code, I tried two demos directly on the Kimi website. The first asked for a voxel art scene, and the model generated it quickly with good visual consistency. The second was a detailed tech stock market research report with visualisation – it handled the data and produced a clear, structured report.
Project Setup with Kilo Code and the Moonshot API
For the real test, I opened a Kids Storybook Generator project in VS Code. It’s a React app that creates illustrated storybooks with voiceover. I configured Kilo Code, an AI coding extension, to connect to the Moonshot API so it would use Kimi K2 Thinking as the backend model.
Getting the Project Running and Adding a Feature
I asked Kimi K2 to understand the project and help me run it. It identified that I needed a Google AI Studio API key for the voice feature and told me exactly where to set it. After adding the key, the project started with npm run dev and generated a storybook successfully. Next, I requested a new feature: a stop/play button for the voiceover. The model edited storybookview.tsx to implement it.
- Configured Kilo Code to use Moonshot API as provider
- Used Kimi K2 to analyse the React project structure
- Set Google AI Studio API key as an environment variable
- Ran npm run dev and confirmed storybook generation
- Requested addition of a stop/play button for voiceover
- Accepted the code changes to storybookview.tsx
Bug Fix and Final Result
After adding the feature, I encountered an API key error when testing the voiceover. Kimi K2 quickly diagnosed the problem – the key hadn't been read correctly in the new context – and resolved it. With that fix, the storybook generation and the new stop/play button worked perfectly.
Conclusion and Final Thoughts
Kimi K2 Thinking delivered on its promises: fast, logical, and remarkably cost-effective. It handled a full React project, understood dependencies, added a feature, and fixed a live error – all while beating Sonnet 4.5 and GPT-5 in these practical scenarios. If you’re into vibe coding, AI development, or automation, this open-source model is well worth a try.
Tools used in this video
- Kimi K2 Thinking
- Moonshot AI (API)
- Kilo Code
- VS Code
- Google AI Studio
- React
- npm
- Kimi.com
Questions
- Is Kimi K2 Thinking open-source?
- Yes, Kimi K2 Thinking is an open-source model from Moonshot AI. You can use it via their API or integrate it into your own setup.
- How does Kimi K2 Thinking compare to Claude Sonnet 4.5 and GPT-5?
- In my tests, Kimi K2 Thinking matched or outperformed both in reasoning, speed, and cost-efficiency, especially in a real React project and web-based tasks like voxel art generation and financial reports.
- Can I use Kimi K2 Thinking with Cursor or Windsurf?
- The model isn’t directly built into those tools, but you can connect it through the Moonshot API with extensions like Kilo Code. The same approach works for any IDE that supports custom API endpoints.
- What error did you encounter and how was it fixed?
- After adding a voiceover stop/play button, the app threw an API key error. Kimi K2 quickly identified that the environment variable wasn’t being read correctly in the new code context and corrected the import, resolving the issue.
Topics
Chapters
- 00:00:00 Introduction to Kimik2 Thinking Model & Cost
- 00:00:11 Benchmarks and Deciding on a Real Project Test
- 00:00:22 Project Introduction: Kids Storybook Generator (VS Code & Kilo Code)
- 00:00:38 Test 1: Voxel Art Scene Generation (Official Website)
- 00:01:17 Test 2: Detailed Tech Stock Market Research Report & Visualization
- 00:02:46 Project Setup in VS Code with Kilo Code
- 00:03:09 Demo of the Existing Kids Storybook Generator App
- 00:04:10 Configuring Kilo Code to Use Moonshot (Kimik2's Parent Company) API
- 00:05:39 Using Kimik2 to Understand and Run the React Project
- 00:06:55 Kimik2 Identifies Required Environment Variables (Google AI Studio API Key)
- 00:08:26 Setting the API Key and Running the Project (npm run dev)
- 00:09:03 Testing the Running Project (Successful Storybook Generation)
- 00:09:44 Adding a New Feature: Stop and Play Button for Voiceover
- 00:10:27 Kimik2 Edits storybookview.tsx to Implement the Feature
- 00:11:06 Testing the New Feature (Encountering an API Key Error)
- 00:11:37 Kimik2 Bug Fixing: Resolving the API Key Error
- 00:12:00 Final Test: Confirming the Storybook Generation and the New Feature
- 00:12:34 Conclusion and Final Thoughts on Kimik2's Performance
Video description and links
This article was written from the video's description with the help of AI. Watch the video for the full walkthrough.



