Tencent Hunyuan 3.0 officially launched, focusing on improving task execution capability.

Nearly half a year after joining Tencent, Tencent’s Chief AI Scientist Yao Shunyu’s first report card—Hy3 official release (Hunyuan 3.0 official release)—has finally been released.

On July 6, Wall Street Insights discovered that Tencent’s AI assistant Yuanbao has launched the Hy3 official release. This is 74 days after Tencent released the Hy3 preview (Hunyuan 3.0 preview).

Hy3 is a MoE language model that integrates fast-and-slow reasoning, with total parameters of 295B, activated parameters of 21B, and a maximum supported context length of 256K. Judging from the model size, Hy3 has not changed compared with the previously released Hy3 preview.

That is to say, Tencent sticks to a Hy3 approach that does not pursue large-scale parameters, positioning it as a “solution that balances performance and cost-effectiveness,” with the goal of becoming one of the best options for practical deployment across most business scenarios.

From Tencent’s perspective, 300B is the optimal balance between capability and efficiency. Capabilities such as complex reasoning, long-context understanding, and instruction following have already been fully unlocked at this scale. Continuing to expand parameter size brings sharply diminishing marginal returns—if the investment doubles, capability improvements are often only in single-digit percentage points.

It is reported that based on the Hy3 preview, the Hy3 official release focuses on optimizing Coding Agent task execution and productivity scenarios, as well as complex reasoning and mathematical abilities, and has fixed several known issues that existed in the Hy3 preview.

Meanwhile, Hy3 pricing has been further reduced. The current price is 1 yuan per million tokens for input, 4 yuan per million tokens for output, and 0.25 yuan per million tokens for input cache hits.

Wall Street Insights learned from Tencent that Hy3 has already been connected to multiple business units under Tencent, including WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima, and more. The API has been launched on Tencent Cloud TokenHub, and multiple overseas API platforms will also be connected in sequence.

Hy3 is Tencent’s calibration of pace in the second half of AI.

In the second half of 2025, Tencent has relatively intensively carried out organizational upgrades of the Hunyuan large-model team and restructured workflows. Meanwhile, in February 2026, it also rebuilt large-model R&D infrastructure, including pre-training and reinforcement learning, and further improved data quality.

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

It is reported that during the reconstruction of the Hunyuan large model, Tencent also established three principles for pursuing practicality: first, emphasizing a systematic capability framework and not advocating “specializing in just one area”; second, ensuring evaluation authenticity and proactively stepping away from public leaderboards that are easy to “game”; third, pursuing cost-effectiveness.

Therefore, Tencent chose to release the Hy3 preview first in order to obtain real feedback from the open-source community and users, helping improve the practicality of the Hy3 official release.

Tencent provides data showing that since its release, the number of users on WorkBuddy who have voluntarily chosen the Hy3 preview has grown by 6 times. In internal evaluations of the WorkBuddy office scenario, Hy3’s task success rate increased from 72% to 90%, and the average time spent was reduced by 34%.

In Yuanbao usage, regarding the hallucination phenomenon that exists in long-form text and AI search scenarios, Hy3 teaches the model to produce reliable outputs under complex evidence through fine-grained data cleaning and training constraints.

It is reported that in evaluations based on Yuanbao’s real logs, Hy3’s common-sense error rate has fallen by half compared with the preview version, and the hallucination rate has decreased by more than half.

Worth noting is that, based on the significantly improved Agent capabilities of Hy3, Yuanbao has also launched Agent functionality in parallel.

In hands-on experience, Wall Street Insights found that in Yuanbao’s task prompts, when requirements to generate the corresponding files—such as PPT, Word, Excel, PDF, HTML, and others—are explicitly stated, Yuanbao’s Agent capability can be triggered. It outputs a final complete file, rather than the previous form of pure conversation.

Risk Warning and Disclaimer

        The market involves risks; investing requires caution. This article does not constitute personal investment advice, and it does not take into account the special investment objectives, financial conditions, or needs of any individual user. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at your own risk, and you shall bear the responsibility.