Video tutorial · The Metaverse Guy
HY3 Preview: Testing Tencent’s Open-Source AI Model for Coding
I put HY3 Preview through real coding projects—web scraping, 3D simulation, debugging—and compared it to Claude, Gemini, and GPT. Here’s what I found.
- Published
- 26 May 2026
- Length
- 8 min
- Views
- 13,815 views
In this video, I test HY3 Preview, Tencent’s newly released open-source AI model, on a range of coding tasks using OpenRouter and KiloCode. I built an AI web scraper for just $0.08, a 3D Rubik’s Cube simulator, and pushed the model on debugging and long-context workflows. The results surprised me, and I compare its performance and cost against Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek V4, and GPT models.
Introduction
I’ve been searching for the cheapest AI tools that can still deliver real value for day-to-day development. When Tencent released HY3 Preview, an open-source model available through OpenRouter, I had to test it properly. In this video, I put it through a series of real coding projects—not just benchmarks—to see if it could genuinely replace more expensive models like Claude or Gemini in my workflow.
What is HY3 Preview?
HY3 Preview is a new open-source AI model from Tencent. It’s multimodal, fast, and built with a long context window, making it suitable for complex coding tasks. Despite its capabilities, it’s flown under the radar for many developers. I wanted to see if it deserved more attention, especially given its pricing on OpenRouter.
Real-World Coding Tests
I ran HY3 Preview through several projects inside KiloCode, using OpenRouter as the API provider. Each test was designed to stress a different aspect of the model: generation, reasoning, debugging, and handling long contexts.
- AI web scraper: I built a tool that scrapes data from any website using a single prompt. The entire project cost just $0.08.
- 3D Rubik’s Cube simulator: I generated a fully interactive 3D Rubik’s Cube using Three.js. The model produced working code quickly.
- Debugging and self-testing: I gave it broken code and asked it to find and fix issues. It identified problems and suggested corrections reliably.
- Long-context coding: I fed it a large codebase and asked for modifications across multiple files. It maintained context well and made accurate changes.
- Multi-step agent tasks: I chained several instructions together to simulate an AI agent workflow. The model followed the sequence without losing track.
Performance and Cost
HY3 Preview impressed me with its speed and reasoning ability. It handled complex logic and generated clean, functional code. The standout feature, though, is the cost. At $0.08 for a full project, it’s one of the cheapest capable models I’ve used. For developers who need to iterate quickly without burning through API credits, this is a serious option.
How It Compares to Other Models
I compared HY3 Preview directly against Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek V4, and GPT models. While the larger models still have an edge on extremely nuanced tasks, HY3 Preview held its own on practical coding work. For many everyday development tasks, the difference was negligible, and the cost savings were significant. It’s a compelling alternative if you’re looking to reduce your AI spend without sacrificing too much capability.
Getting Started with HY3 Preview
You can try HY3 Preview right now via OpenRouter. I used it inside KiloCode, but any tool that supports OpenRouter’s API will work. Just select the model from the list, and you’re ready to start building. I’ve included the setup steps in the video if you need a walkthrough.
Final Thoughts
HY3 Preview might be the most underrated open-source model released this year. It’s cheap, fast, and surprisingly capable for real-world development. I’m not saying it will replace Claude or Gemini for everyone, but for a lot of daily coding tasks, it’s absolutely worth testing. I’d love to hear your thoughts in the comments—would you switch to HY3 Preview for your own projects?
Tools used in this video
- HY3 Preview
- OpenRouter
- KiloCode
- Three.js
Questions
- How much does HY3 Preview cost to use?
- In my test, building a full AI web scraper cost only $0.08. Pricing is through OpenRouter and is extremely low compared to models like Claude Opus or Gemini.
- Can HY3 Preview handle long code files?
- Yes, it supports a long context window. I tested it with a large codebase and multi-file modifications, and it maintained context without issues.
- Is HY3 Preview good for debugging?
- I gave it broken code and it successfully identified and fixed the problems. Its reasoning ability makes it a solid choice for debugging tasks.
- How do I access HY3 Preview?
- You can access it through OpenRouter. I used it inside KiloCode, but any OpenRouter-compatible tool will work. Just select HY3 Preview from the model list.
Chapters
- 00:00 Intro
- 00:08 AI Web Scraper Built for $0.08
- 00:24 What is HY3 Preview?
- 00:42 Why I Started Testing This Model
- 00:54 3D Rubik’s Cube Demo
- 01:24 Why Most People Missed HY3 Preview
- 01:47 HY3 Preview vs Claude, Gemini & GPT
- 02:18 OpenRouter Ranking & Benchmarks
- 02:45 How to Use HY3 Preview in KiloCode
- 03:10 Real Coding Test Results
- 04:12 Reasoning Performance Breakdown
- 04:44 Long Context Window Test
- 05:20 Debugging & Self-Testing With Chrome DevTools
- 06:00 HY3 Preview for AI Coding Agents
- 06:42 Why This Matters for Open-Source AI
- 07:20 Final Thoughts on HY3 Preview
- 07:45 Outro
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.



