<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/18afb992-dd79-4b3a-9b6f-410e99a95a3d.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/18afb992-dd79-4b3a-9b6f-410e99a95a3d.png"/>
Author | Huang Yu
On April 23, Tencent officially unveiled and open-sourced Hy3 preview (Hunyuan 3.0 Preview).
Hy3 preview is a MoE language model integrating fast and slow thinking, with 295 billion total parameters, 21 billion activated parameters, and support for a maximum context length of 256K.
It is evident that Hy3 preview does not pursue massive parameter scales but positions itself as &#34;balancing performance and cost-effectiveness,&#34; aiming to become one of the optimal choices for implementation in most business scenarios.
According to Tencent, 300B represents the optimal balance between capability and efficiency. Capabilities such as complex reasoning, long-context understanding, and instruction following are fully unleashed at this scale; continuing to expand the parameter size yields significantly diminishing marginal returns—doubling investment often results in capability improvements of only single-digit percentage points.
In addition to daily conversation capabilities (casual chat, writing, search, etc.), Hy3 preview also focuses on enhancing skills in Coding, agents, instruction following, and context understanding. It has already been deployed in numerous internal Tencent products including Yuanbao, ima, WorkBuddy, and CodeBuddy.
Hy3 represents a rhythm calibration for Tencent in the second half of the AI race.
Over the past few months, Tencent has conducted intensive organizational upgrades and workflow restructuring for its Hunyuan large model team. In February this year, it also re-established large model R&amp;D infrastructure covering pre-training and reinforcement learning, while further improving data quality.
At that time, Tencent also established three principles for pursuing model practicality: first, emphasizing systematic capability development rather than promoting &#34;specialization&#34; in narrow areas; second, ensuring evaluation authenticity by actively moving away from public leaderboards prone to &#34;gaming&#34;; third, pursuing cost-effectiveness.
Hy3 preview is not only the first large model after Hunyuan underwent a full-link reconstruction but also Yao Shunyu's first report card since joining Tencent as Chief AI Scientist, Head of AI Infrastructure Department, and Head of Large Language Model Department.
According to Wall Street Insights, training for Hy3 preview commenced at the end of January 2026, taking less than three months from training to launch. This is regarded internally by Tencent as the beginning of the Hunyuan large language model attempting to solve real-world problems.
Yao Shunyu stated that Hy3 preview is the first step in reconstructing the Hunyuan large model. Tencent hopes that this open-source release will garner authentic feedback from the open-source community and users to help improve the practicality of the official Hy3 version.
Simultaneously, &#34;we are also continuing to expand the scale of pre-training and reinforcement learning to raise the upper limit of model intelligence, and through deep Co-Design with numerous Tencent products, continuously enhance the model's comprehensive performance in real-world scenarios, while beginning to explore distinctive model capabilities,&#34; said Yao Shunyu.
It is reported that during the R&amp;D of Hy3 preview, the Hunyuan model team conducted co-design with the Yuanbao product team.
The Hunyuan team believes that model evaluation should not be a simple stacking of leaderboards but an adaptation to complex capability systems and implementation in actual business scenarios. Therefore, on one hand, the team built over 50 benchmarks to assess the model's actual capabilities and deployability; on the other hand, they closely aligned with internal Tencent businesses to enable the model to learn and evolve through practical applications.
The launch of Hy3 preview is also an important signal of accelerated evolution in Hunyuan R&amp;D. According to Wall Street Insights, under the support of new infrastructure and technical concepts, larger-sized models within Hunyuan are already in development.
Now that the competition in AI technology has entered the second half, the collaborative effect of large models within a complete workflow, or their ability to &#34;execute tasks,&#34; has become the focus of competition. This is precisely why Hy3 preview focuses on enhancing Coding, intelligence, instruction following, and context learning capabilities.
To verify Hy3 preview's operational capabilities, the Hunyuan model team conducted manual evaluations for internal users, covering typical usage environments such as coding and general workflows. Data provided by Tencent shows that Hy3 preview achieved an overall win rate of approximately 55%–56% in blind user evaluations.
Currently, Hy3 preview has also been integrated into internal Tencent AI Agent products such as CodeBuddy and WorkBuddy.
Data provided by Tencent indicates that on CodeBuddy and WorkBuddy products, Hy3 preview reduced first token latency by 54%, decreased end-to-end duration by 47%, and increased success rates to over 99.99%.
In actual user environments, Hy3 preview has stably driven complex agent workflows up to 495 steps, covering diverse office scenarios such as document processing, data analysis, knowledge retrieval, and MCP toolchain orchestration.
Tencent Senior Executive Vice President and CEO of Cloud and Smart Industries Group, Tang Daoxing, publicly stated in March that the current application paradigm of artificial intelligence is shifting from &#34;Chatbot&#34; to &#34;AI Agent.&#34; AI implementation is not merely an algorithmic challenge but also an engineering one—as the capability gap among mainstream large models narrows, enterprises are no longer competing on &#34;whose model is stronger,&#34; but rather on who can effectively utilize models through engineering means.
Clearly, Tencent is attempting to prove that even if the model itself is not the absolute top-tier, as long as the &#34;foundation&#34; is stable, interfaces are abundant, and engineering capabilities are strong, it can still win the ecosystem battle in the Agent era.
The release of Hy3 preview marks that Tencent is no longer obsessed with the myth of stacking parameter scales but chooses to leverage Tencent's vast social and tool ecosystems for high-efficiency &#34;technology cultivation through combat&#34; based on a 300B parameter baseline.
How far this sense of rhythm allows Tencent to go in the second half of the Agent race will depend on whether the official Hy3 version can truly achieve a qualitative transformation from &#34;reading ten thousand books&#34; to &#34;traveling ten thousand miles.&#34;

TCEHY

TCTZF

00700

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

- On April 23, Tencent unveiled the Hy3 preview, a MoE language model with 295 billion parameters, emphasizing performance and cost-effectiveness.  
- The model's capabilities include improved coding, context understanding, and instruction adherence, and it has been integrated into several internal Tencent products.  
- Hy3 preview signifies Tencent's shift towards practical AI applications and aims to gather community feedback for its improvement, marking a new phase in its AI development strategy.

Tencent Hy3 Preview Launches, Report Card Released After Yao Shunyu's Join