Hy3 preview: A Rebuilt Hunyuan, a 21B-Active MoE, and a New Reasoning Receipe
1. What is Hy3 Preview?
Hy3 preview is an open-source fusion reasoning model that integrates fast and slow thinking. With 295B total parameters and only 21B activated parameters, it stands out as the most efficient model in terms of activation.
Currently, there are two mainstream paths in global LLM development: one pursues higher intelligence through extremely large total parameters and high activation parameters; the other seeks cost-effectiveness by using large total parameters but small activation parameters. Hy3 preview carves its own unique path—achieving performance close to the second approach with even smaller total and activated parameters, through innovations in architecture, data quality, and training efficiency.
Within its 295B total parameters, every expert is meticulously optimized in terms of training quality, routing precision, and data mixture, ensuring that every activated parameter delivers maximum utility. In terms of capabilities, Hy3 preview delivers comprehensively improved complex reasoning and significantly enhanced coding performance. Its distinctive features include Context Learning, which excels in contextual memory and understanding.
2. Inside Tencent's product ecosystem
With the integration of Hy3 preview, Tencent Yuanbao is capable not only of general-purpose functions such as casual conversation and writing but also of accurately understanding users' core needs. By leveraging deep search to gather information, it integrates detailed and actionable plans for users, with notable improvements in both content professionalism and structural organization. Furthermore, Hy3 preview can perform tool calls based on user prompts to retrieve the latest literature, news, and other real-time information.
For solution-oriented questions, Hy3 preview can also provide real-time responses.
In complex reasoning (i.e., challenging problems), Hy3 preview demonstrates the reasoning capabilities that reflect a model's true intelligence level, with the fastest inference speed and highly clear step-by-step planning.
The demonstrated efficiency stems from Tencent Hunyuan’s continuous in-house innovation in MoE architecture. Unlike traditional MoE where all experts are the same size, Hunyuan introduces a differentiated expert size design, routing tokens of varying difficulty to experts with different capacities. Experimental results show that under the same computational budget, this approach significantly outperforms traditional MoE. Additionally, Hunyuan incorporates a P-Penalty Loss, which penalizes the model’s tendency to favor large experts and encourages it to activate more small experts, thereby improving computational efficiency without sacrificing performance.
What others achieve with 37B–40B activated parameters, Hunyuan accomplishes with just 21B. The precision of routing determines the true value of these 21B activated parameters — this is a core algorithmic moat that cannot be overtaken simply by scaling up hardware or parameters.
3. Benchmarks & Case Studies
In knowledge extraction and interactive learning, such as turning a textbook into an interactive quiz game, the model must extract structured knowledge from long documents and generate a complex interactive frontend. Specifically, it reads a technical manual of over 100 pages, automatically designs 10 in-depth test questions, and packages them into a single-file HTML game with scoring and answer explanation features. This process leverages Tencent’s self-developed AI Agent, WorkBuddy, which integrates Hy3 preview, and tests the model’s capabilities in long-context understanding, multi-instruction following, and code generation.
In long-text understanding and code generation, the model processes information quickly and excels at comprehending complex instructions. Its task planning capability is outstanding, enabling it to break down complex tasks into clear, logical steps and execute them in an orderly manner. In terms of text extraction, the model accurately captures key information from lengthy documents with very high precision. Built on these understanding and planning abilities, the code it generates is bug-free, runs correctly the first time, and executes successfully, demonstrating stable and reliable overall performance.
Compared with other models, Hy3 preview stands out in three key areas: long-text understanding and summarization, tool usage, and code generation. When dealing with lengthy documents, it doesn't just capture the core information and logical flow—it also connects key points scattered across different sections and produces clear, well-structured summaries that highlight what matters most. In tool usage, it accurately grasps what the user wants to do, decides on its own when to call a tool and which one to use, and properly processes the returned results to complete complex tasks like retrieval, calculation, and querying. As for code generation, it quickly writes runnable code from natural language descriptions—clean logic, few syntax errors, and often works on the first try. These capabilities work together, making Hy3 preview much more effective at handling complex, real-world tasks.
4. Takeways
Hy3 preview is the first model trained after the rebuild of Hunyuan's large model architecture. From architectural upgrades and infrastructure reconstruction to official release, it took less than three months. The model delivers significant and comprehensive improvements in areas such as chat, code generation, Agent capabilities, mathematical reasoning, and long-context understanding. It has already been integrated into Tencent's internal products, including Yuanbao, CodeBuddy, and WorkBuddy. Rather than blindly pursuing larger parameter counts, the Hunyuan team places greater emphasis on data quality and architectural innovation. Through Co-Design, the model is deeply integrated with Tencent's core businesses—social networking, gaming, and advertising—allowing it to learn and evolve in real-world scenarios. Meanwhile, Tencent has accelerated its efforts in the Agent domain, developing products such as WorkBuddy in-house, and leveraging a cloud-native "sandbox" infrastructure at the bottom layer to support hundreds of thousands of concurrent requests. This has gradually formed a systematic advantage that combines model capability, technical strength, product experience, and ecosystem synergy.
