From Preview to Production: Tencent's Hy3 Model Has Arrived
Free in Kilo for a limited time
Remember Hy3 Preview—the model that completely shook up our workflows during its trial phase? As one of our most popular free preview models for coding, it quickly became the go-to assistant for devs looking to build smarter, faster, and more efficiently.
We’re thrilled to announce that the much-anticipated General Availability (GA) release of Hy3, totally free for a limited time everywhere you use Kilo. That means you get elite, production-grade AI assistance right in your VS Code sidebar, cloud agents, our CLI, you name it! All completely free of charge with high limits thanks to compute from Novita.
Why Hy3 is Your New Coding Superpower
Developed by the team at Tencent, the GA release of Hy3 isn’t just a minor update; it’s a massive leap forward in reasoning, agentic workflows, and production-grade stability.
Moving beyond the preview stage, the team scaled up Reinforcement Learning (RL) training and fine-tuned post-training data quality. In blind tests with real-world expert workflows, Hy3 consistently outperformed similar-sized models and rivaled flagship open-source models with 2x to 5x the parameters.
By utilizing Kilo’s VS Code extension, you can seamlessly bring these massive upgrades straight into your active workspace. Key real-world improvements in the GA release include:
Production-Grade Tool Calling: Massive upgrades to tool-call success rates and error recovery, drastically minimizing invalid calls or infinite loops while using agentic features in your editor.
Plummeting Hallucinations: Thanks to rigorous data cleaning and specific training constraints, the internal hallucination rate has dropped from 12.5% to just 5.4%—meaning more accurate code syntax and fewer phantom bugs.
Long-Context Intent Tracking: Upgraded multi-turn dialogue capabilities ensure your complex coding intents don’t drift or decay over long, multi-step refactoring sessions in your codebase.
In my view, Tencent’s expertise in video game development across their massive portfolio has helped improve this model’s capabilities around coding in general. It’s highly efficient (following the trend from both closed and open labs lately) and strong at both front-end and back-end development tasks.
Under the Hood: The Specs
Hy3 continues the trend of recent open-weight releases focused on both efficiency and capabilities for daily agentic work. Here are the specs:
Architecture: Mixture-of-Experts (MoE)
Total Parameters: 295B
Activated Parameters: 21B
Context Length: 256K
Number of Experts: 192 experts, top-8 activated
It’s great to see the 256K context length for Kilo users with large codebases and complex projects, and with 295B total parameters (while keeping inference lightning-fast by activating 21B parameters per token), Hy3 is perfectly engineered to parse entire files and tackle massive documentation without breaking a sweat—making it the ultimate companion for your VS Code environment.
We’ll get the stats live in PinchBench and KiloBench soon as well. Watch the Kilo Leaderboard for updates.
Start Building for Free Today
It’s time to move your daily dev projects out of the preview era and into full, robust production without ever leaving your editor.
Just select Tencent Hy3 from the model dropdown wherever you use Kilo. Happy coding!