What Tencent's Hunyuan Hy3 Release Means for Open-Source AI

Tencent has rolled out the full production version of Hunyuan Hy3, its next-generation large language model, marking a major step forward from the preview build the company introduced in April. The release cements Tencent's push to become a serious player in the open-source AI race, offering developers a model that is both powerful and free to build on under the permissive Apache 2.0 license.

Model Architecture: Big Ambitions, Smaller Footprint

Hy3 is built as a Mixture-of-Experts (MoE) model with 295 billion total parameters, but only 21 billion of those parameters are activated during inference. That design lets Tencent keep computational costs down while still delivering performance that the company says matches or beats flagship open-source rivals carrying two to five times as many parameters.

The model also supports an expansive 256K token context window, giving it the capacity to process long documents, extended conversations, or large codebases in a single pass.

Hallucination Rate Cut in Half

One of the standout improvements in the full release is a sharp drop in hallucination rate. Compared to the April preview, Hy3's hallucination rate has been reduced by more than half, now sitting at 5.4%. Alongside that gain in reliability, the model reaches a 90% agent task resolution rate, reflecting a strong ability to complete multi-step, agent-driven tasks accurately.

Pricing and Where to Access Hy3

Tencent has also sharpened Hy3's pricing compared to the preview version. Input tokens now cost RMB 1 per million, down from the preview's RMB 1.2 per million, while output tokens are priced at RMB 4 per million. Developers can access the model through Tencent Cloud's TokenHub platform, as well as on GitHub and Hugging Face, making it broadly available to the open-source developer community.

From a Rapid Preview to a Refined Full Release

Hy3's development timeline moved quickly. Tencent restructured its Hunyuan team in late 2025 under chief AI scientist Yao Shunyu, and the newly rebuilt infrastructure produced the Hy3 preview in under three months, with that preview launching on April 23.

Preview Version Topped OpenRouter's Usage Charts

The preview build didn't just launch fast — it caught on fast. By early May, it had climbed to the top of OpenRouter's weekly call volume chart, logging 3.66 trillion tokens of usage on the aggregation platform. That early traction set the stage for the full release, which builds on the preview's momentum with improved post-training data quality and expanded reinforcement learning computation.

Benchmark Performance Against GLM-5.2

According to reporting, Hy3 outperforms Zhipu's GLM-5.2 across most benchmarks despite operating at roughly half the size. That combination of smaller scale and stronger benchmark results is central to Tencent's pitch that Hy3 delivers flagship-level capability without the resource overhead typically required to achieve it.

Integration Across Tencent's Product Ecosystem

Hy3 isn't limited to standalone developer access — it's already been woven into several of Tencent's own products, including WorkBuddy, CodeBuddy, Yuanbao, and QQ. That kind of built-in distribution gives Hy3 an installed user base well beyond what a standalone model release would typically reach. The announcement also coincided with a market reaction: Tencent's Hong Kong-listed shares rose more than 4% on the day of the release.

Intensifying Competition Among Chinese AI Labs

Hy3's launch adds fuel to an already heated race among Chinese AI developers. Open-source models from DeepSeek, Alibaba's Qwen, and Zhipu have all been competing hard for developer adoption, and Tencent's aggressive pricing combined with the permissive Apache 2.0 license signals a clear strategy: prioritize accessibility and ecosystem integration in hopes that Hy3 becomes a default infrastructure layer for enterprise AI across China.