Skip to main content

Tencent Hunyuan Hy3 Launches: Smarter, Cheaper, and More Practical

On July 6, Tencent officially released Hunyuan Hy3, the latest version of its large language model. Compared to the preview version that came out in April, Hy3 shows a significant boost in intelligence—often matching or even beating models that are two to five times larger. And here's the kicker: it's also cheaper. The API pricing is set at 1 yuan per million tokens for input and 4 yuan for output, with a cache hit rate bringing input down to just 0.25 yuan.

Hy3 uses a Mixture-of-Experts (MoE) architecture, with a total of 295 billion parameters but only 21 billion activated per token. That means it's efficient without sacrificing smarts. It can handle up to 256,000 tokens of context, which is plenty for long documents or complex conversations.

Image

A Smarter Model for Real-World Tasks

What makes Hy3 stand out is how much better it has gotten at practical tasks. In a blind test with 270 experts working on real job scenarios, Hy3 scored an average of 2.67 out of 4, beating GLM-5.1's 2.51. The biggest wins came in areas like front-end development, data storage, and CI/CD pipelines. That's not just a lab result—it's a sign that the model is ready for the real world.

Tencent has already rolled Hy3 into several of its products. WorkBuddy, the company's AI office assistant, saw its task success rate jump from 72% to 90% after switching to Hy3, while average task time dropped by 34%. Users seem to agree: the number of people choosing Hy3 on WorkBuddy has grown sixfold since the preview launched.

Image

Cutting Down on Hallucinations

One of the biggest headaches with AI models is when they make stuff up—hallucinations. Hy3 tackles this head-on. In Yuanbao, Tencent's AI assistant, the model's common sense error rate was cut in half compared to the preview, and hallucinations dropped by more than half. That's thanks to better training data and stricter constraints during fine-tuning.

Yuanbao now uses Hy3 to power its agent functions. You can just tell it what you need—a PPT, a Word doc, an Excel sheet—and it'll handle the whole thing. And it's free. In internal tests, Hy3 outperformed several other top Chinese models in both office and daily life scenarios.

Image

More Than Just Talk

Hy3 isn't just about chat. It's also being used in WeChat's AI avatars and customer service, where intent recognition accuracy hit 98.94%. Even when users' expressions are incomplete, the model can figure out what they mean without jumping to conclusions. In WeChat Reading, tag annotation accuracy improved by 14.1% over the preview, and classification efficiency went up by 8.4%.

For gamers, the AI assistant in "Path of Exile: Arrival" on WeGame saw its multi-round reasoning success rate climb to 92%, while hallucinations dropped from 4.5% to 2.8%. That means fewer weird answers and a smoother experience.

Open and Accessible

Tencent is keeping Hy3 open. It's released under the Apache 2.0 license, which means developers can use it for commercial projects without worrying about fees. The model is already available on Tencent Cloud TokenHub, and it's coming to overseas platforms like OpenRouter, Hugging Face, and ModelScope soon.

From the infrastructure overhaul in early 2026 to the preview in April and now the full release, Tencent has moved fast. The company says it will keep pushing the limits of what AI can do, focusing on making these powerful models useful in everyday work and life.

Key Points

  • Hy3 uses MoE architecture: 295B total parameters, 21B activated, 256K context length.
  • Outperforms larger models: Matches or beats models 2-5x its size in many tasks.
  • Real-world improvements: Task success rate up to 90% in WorkBuddy, hallucinations cut by half.
  • Integrated into Tencent products: WorkBuddy, Yuanbao, WeChat, games, and more.
  • Affordable pricing: 1 yuan per million input tokens, 4 yuan output, with cache discounts.
  • Open source under Apache 2.0: Free for commercial use, available on multiple platforms.