Tencent Hy3 preview goes online; Yao Shunyu's report card released after joining

```

Author | Huang Yu

On April 23, Tencent officially unveiled the Hy3 preview (Hunyuan 3.0 Preview Version) and open-sourced it.

Hy3 preview is an MoE language model that blends fast and slow thinking, with a total of 295 billion parameters, 21 billion activated parameters, and supports a maximum context length of 256K.

It’s not hard to see that Hy3 preview does not pursue massive parameter counts, but is positioned at "balancing performance and cost-effectiveness," aiming to become one of the optimal choices for practical application in most business scenarios.

According to Tencent, 300B is the optimal balance point between capability and efficiency. Complex reasoning, long context understanding, and instruction following abilities are fully unleashed at this scale, and the marginal benefit of increasing parameter size declines significantly—doubling investment often only yields single-digit percentage improvements.

Reportedly, besides capabilities for everyday dialogue (chatting, writing, searching, etc.), Hy3 preview also focuses on enhancing abilities in coding, agents, instruction following, and context understanding. It has already been launched in many internal Tencent products such as Yuanbao, ima, WorkBuddy, and CodeBuddy.

Hy3 is a rhythm calibration for Tencent in the second half of the AI race.

In recent months, Tencent has intensively upgraded the organization and workflows of the Hunyuan large model team, and rebuilt the basic infrastructure for large model R&D—including pretraining and reinforcement learning—and further improved data quality.

At that time, Tencent also established three principles for practical model pursuit: First, emphasize systematic capabilities and not favor "subject bias"; Second, ensure authenticity in evaluation and proactively step away from public leaderboards easily "gamed"; Third, pursue cost-effectiveness.

Hy3 preview is not only the first large model after rebuilding the entire Hunyuan chain but also the first report card since Yao Shunyu, Tencent’s Chief AI Scientist and head of AI Infra and Large Language Model departments, joined Tencent.

According to Wallstreetcn, Hy3 preview started training at the end of January 2026, and it took less than three months from training to launch. Tencent internally views this as the starting point of using Hunyuan large language models to tackle real-world problems.

Yao Shunyu stated that Hy3 preview is the first step in Hunyuan large model reconstruction. Tencent hopes that by open-sourcing and releasing it, the feedback from the open-source community and users will help improve the practicality of the official Hy3 release.

At the same time, "We are also continuing to expand pretraining and reinforcement learning scale, increase the model’s intelligence ceiling, and through deep co-design with many Tencent products, continually improve overall performance in real scenarios and begin exploring specialized model capabilities," said Yao Shunyu.

It is reported that during Hy3 preview's R&D process, the Hunyuan model team and Yuanbao product team conducted co-design.

The Hunyuan team believes that model evaluation is not just simple leaderboard stacking, but adaptation to complex capability systems and real-world business scenarios. Therefore, the team built more than 50 benchmarks to assess actual capability and applicability, and closely aligned with internal Tencent businesses, allowing the model to learn and evolve in real applications.

The launch and release of Hy3 preview is also an important signal of accelerated Hunyuan R&D evolution. Wallstreetcn learned that with new infrastructure and technical concepts, a larger Hunyuan model is already on its way.

With AI technology competition entering its second half, the collaborative effect of large models across complete workflows, or their "task execution" capability, has become the focus. This is why Hy3 preview emphasizes enhancing coding, intelligence, instruction following, and context learning.

To verify Hy3 preview's working ability, the Hunyuan model team conducted manual evaluation with internal users, covering typical environments such as coding and general workflows. Tencent’s data shows that Hy3 preview achieved an overall win rate of about 55%–56% in user blind evaluations.

Currently, Hy3 preview has been integrated into Tencent's internal AI Agent products like CodeBuddy and WorkBuddy.

Tencent’s data shows that on CodeBuddy and WorkBuddy, Hy3 preview’s first token latency decreased by 54%, end-to-end duration reduced by 47%, and the success rate increased to 99.99%+.

In actual user environments, Hy3 preview has stably powered complex Agent workflows up to 495 steps, covering scenarios such as document processing, data analysis, knowledge retrieval, MCP toolchain orchestration, and other diversified office situations.

Tencent Senior Executive Vice President and CEO of the Cloud and Smart Industry Group, Tang Daosheng, stated publicly in March that the AI application paradigm is shifting from "Chatbot" to "AI Agent." AI implementation is not just an algorithm problem, but an engineering problem. As the capability gap of mainstream large models narrows, companies compete not on "whose model is better," but on who can use the model best through engineering means.

Clearly, Tencent is trying to prove that even if the model itself is not top-tier, as long as the "chassis" is stable, there are enough interfaces, and strong engineering capabilities, it can still win the ecological battle in the Agent era.

The release of Hy3 preview signifies that Tencent is still not obsessed with the myth of scale, but chooses to use Tencent’s massive social and tool ecosystem for highly efficient "skill cultivation through battle" at the 300B parameter baseline.

How far this sense of rhythm can carry Tencent in the second half of the Agent era depends on whether the official Hy3 version can truly achieve a qualitative leap, after "reading ten thousand books" and "traveling ten thousand miles".

Risk Warning and DisclaimerThe market is risky; investment should be cautious. This article does not constitute personal investment advice and does not consider individual user’s specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Investment is at one's own risk. ```