Skip to main content

Tencent's Hunyuan Hy3: Smarter, Cheaper, and Ready for Work

On July 6, Tencent officially released Hunyuan Hy3, the latest version of its large language model. Compared to the preview that came out in April, Hy3 brings a noticeable jump in intelligence—especially in complex reasoning, following instructions, and handling agent tasks. And here's the kicker: it does all this while being smaller and cheaper than many flagship models.

What's Under the Hood?

Hy3 uses a Mixture-of-Experts (MoE) architecture with 295 billion total parameters, but only 21 billion are active at any time. That means it's efficient without sacrificing smarts. It supports a context length of up to 256K tokens—enough to chew through long documents or multi-turn conversations.

The preview version already showed a big improvement over Hy2, especially in coding, office productivity, and agent capabilities. Now, with better training data and more compute, Hy3 has pushed even further. In blind tests with 270 experts, Hy3 scored higher than GLM-5.1 in real work scenarios—particularly in front-end development, data storage, and CI/CD tasks.

Real-World Results

Tencent didn't just run benchmarks; they put Hy3 to work across their own products. Since the preview launched, daily token consumption has shot up 20 times—a sign that developers and businesses are finding it genuinely useful.

Take WorkBuddy, China's most popular AI office assistant. Users who chose the Hy3 preview grew sixfold. In internal tests, task success rates jumped from 72% to 90%, and average completion time dropped by 34%. That's not just a number—it means fewer headaches and faster workflows.

Yuanbao, Tencent's AI chatbot, also saw big gains. The common sense error rate halved, and hallucinations dropped by more than half. Thanks to Hy3's improved agent skills, Yuanbao can now handle complex requests—like creating a PPT or editing a spreadsheet—directly from a chat. And it's free.

Image

Other products like ima and Marvis also reported major improvements. In agent tasks, system stability hit 95.1%. For Marvis, task completion rates reached 93.7%, up 12.7% from the preview. And when multiple agents worked together, correct dispatch rates climbed to 92%.

Even WeChat and gaming got a boost. In WeChat Official Accounts, intent recognition accuracy hit 98.94%. In the game "Path of Exile: Arrival," the AI assistant's success rate for multi-step reasoning improved to 92%, while hallucinations dropped from 4.5% to 2.8%.

Image

Pricing and Open Source

Tencent is keeping Hy3 affordable. Input costs 1 yuan per million tokens, output is 4 yuan, and cached input is just 0.25 yuan. That's competitive, especially given the performance.

And it's open source under Apache 2.0, which means developers can use it commercially without worry. Hy3 is already available on Tencent Cloud TokenHub, and it's rolling out to platforms like Hugging Face, ModelScope, OpenRouter, and more.

Image

What's Next?

Tencent says this is just the beginning. They've rebuilt the model from the ground up in less than six months, and they're already planning more updates. The goal is to keep pushing the limits of what a smaller, efficient model can do—and make sure it works in real life, not just in labs.

Key Points

  • Hy3 is smarter: Outperforms many larger models in reasoning, coding, and agent tasks.
  • Real-world tested: Integrated into WorkBuddy, Yuanbao, WeChat, and games with measurable improvements.
  • Cost-effective: Pricing starts at 1 yuan per million tokens; open source under Apache 2.0.
  • Available now: API on Tencent Cloud, plus multiple overseas platforms coming soon.