After nearly six months at Tencent, the first major achievement from Chief AI Scientist Yao Shunyu has been released: the official version of Hy3, also known as Hunyuan 3.0.

On July 6th, it was observed that Tencent's AI assistant Yuanbao has integrated the official Hy3 version. This launch comes 74 days after the company initially released the Hy3 preview model, or Hunyuan 3.0 preview.

Hy3 is a mixture-of-experts (MoE) language model that combines fast and slow thinking processes. It has a total of 295 billion parameters, with 21 billion activated parameters, and supports a maximum context length of 256K. In terms of these core specifications, the official Hy3 remains unchanged from the previously released preview version.

This indicates that Tencent is maintaining its strategy with Hy3, avoiding a pursuit of massive parameter counts. The model is positioned to offer a balance of "performance and cost-effectiveness," with the goal of becoming one of the optimal choices for practical implementation across most business scenarios.

From Tencent's perspective, the 300-billion-parameter scale represents the optimal balance between capability and efficiency. Complex reasoning, long-context understanding, and instruction-following abilities are fully realized at this level, while the marginal gains from further increasing the parameter scale diminish significantly—often resulting in only single-digit percentage improvements despite a doubling of investment.

It is reported that, building upon the Hy3 preview, the official version has been optimized for Coding Agent task execution, productivity scenarios, complex reasoning, and mathematical capabilities. It also addresses several known issues that were present in the preview model.

Concurrently, the pricing for Hy3 has been further reduced. The current rates are 1 yuan per million tokens for input, 4 yuan per million tokens for output, and 0.25 yuan per million tokens for cached input hits.

Information obtained indicates that Hy3 is already integrated into several of Tencent's internal services, including WorkBuddy/CodeBuddy, Yuanbao, Marvis, and ima. Its API is now available on Tencent Cloud's TokenHub, with plans for gradual integration across multiple overseas API platforms.

The release of Hy3 represents a strategic recalibration for Tencent in the current phase of AI development.

Since the latter half of 2025, Tencent has undertaken a series of intensive organizational upgrades and workflow restructuring within its Hunyuan large model team. Furthermore, in February 2026, the company re-established its foundational infrastructure for large model development, encompassing pre-training and reinforcement learning, while also taking steps to further enhance data quality.

Yao Shunyu has previously stated that the Hy3 preview was the first step in rebuilding the Hunyuan large model.

It is understood that during this rebuilding process, Tencent established three guiding principles focused on model practicality: first, emphasizing a comprehensive and balanced capability system, avoiding over-specialization in narrow areas; second, prioritizing authentic evaluation by moving beyond public leaderboards that are susceptible to manipulation; and third, consistently pursuing cost-effectiveness.

Consequently, Tencent chose to release the Hy3 preview first to gather genuine feedback from the open-source community and users, which would help enhance the practicality of the official Hy3 release.

Data provided by Tencent shows that since its release, the number of users on WorkBuddy who independently chose the Hy3 preview has increased sixfold. In internal evaluations of Hy3 within WorkBuddy's office scenarios, task success rates improved from 72% to 90%, while the average time required for tasks was reduced by 34%.

In applications for the Yuanbao assistant, Hy3 tackles hallucination issues in long-text and AI search scenarios through fine-grained data cleaning and training constraints, teaching the model to produce reliable outputs even when dealing with complex evidence.

It is reported that in evaluations based on real logs from Yuanbao, the rate of common-sense errors in Hy3 was halved compared to the preview version, and the hallucination rate decreased by more than half.

Notably, leveraging the significantly enhanced Agent capabilities of Hy3, the Yuanbao assistant has simultaneously launched its own Agent functionality.

In practical testing, it was found that when a user's task prompt explicitly requests the generation of specific file types—such as PPT, Word, Excel, PDF, or HTML—this triggers Yuanbao's Agent capability. The assistant then outputs a complete, finalized file, moving beyond the previous model of purely conversational responses.