On July 6, 2026, Tencent officially launched the Hunyuan Hy3 large language model, marking a pivotal turning point for the internet giant in the AI arena. From Pony Ma's public admission at the start of the year that they were "moving too slowly" to launching a flagship model with a 90% Agent task success rate just six months later, Tencent Hunyuan has completed a comprehensive transformation from infrastructure rebuild to product deployment.

Executive-Led Revolution: Infrastructure Rebuilt in Six Months

At Tencent's annual meeting in January 2026, Pony Ma made a rare public criticism of the company's AI operations for "moving too slowly," directly attributing the problem to insufficient infrastructure and low model iteration frequency. This statement quickly triggered deep organizational restructuring.

President Martin Lau promptly announced a restructuring of the Hunyuan team's organizational structure and workflows, focusing on improving data quality and comprehensively rebuilding pre-training and reinforcement learning infrastructure. Shunyu Yao, Tencent's Chief AI Scientist recruited from OpenAI, became the key figure in this transformation. His first move upon taking office was to "completely rebuild pre-training, reinforcement learning, and infrastructure."

The timeline clearly demonstrates Tencent's execution capability: infrastructure reconstruction began in late January, the Hy3 preview version was released and open-sourced on April 23, and the official version launched on July 6. The complete development chain from infrastructure rebuild to product feedback took less than six months.

Market response validated the strategic pivot. As of June 30, Tencent's token call share on OpenRouter rose to 8.7%. On the internal office platform WorkBuddy, the number of users actively choosing Hy3 preview increased sixfold, while daily token consumption surged twentyfold.

Technical Foundation: A 295B Parameter "Mid-Size Flagship"

Hy3 adopts a seemingly contradictory design philosophy: total parameters reach 295B (295 billion), but activated parameters are only 21B (21 billion). This MoE (Mixture of Experts) architecture combined with a "fast-slow thinking fusion" mechanism allows the model to maintain high performance while dramatically reducing inference costs.

On the technical specifications front, Hy3 supports an ultra-long 256K context window and adopts the Apache 2.0 open-source license, making it friendly for commercial applications. During the preview version development phase, the team achieved a 40% improvement in inference efficiency through comprehensive optimization of operator performance and quantization algorithms. The official version further enhanced training data quality and reinforcement learning capabilities, achieving the goal of "performance comparable to larger-scale flagship models."

This "mid-size parameter scale + MoE architecture + reinforcement learning" technical approach represents Tencent's balance point between pragmatism and performance.

Agent Capability Leap: 90% Task Success Rate

The significant improvement in Agent capabilities is Hy3's most eye-catching highlight. In WorkBuddy office scenarios, the task success rate jumped from 72% to 90%, an 18 percentage point increase, while average processing time decreased by 34%. This dual improvement means users can not only complete more tasks but also get results faster.

Real-world testing data from multiple platforms comprehensively validates these capability improvements:

  • ima platform: System stability reached 95.1%, knowledge base reasoning quality improved by approximately 19%
  • Yuanbao app: Hallucination rate decreased by approximately 44% compared to preview version, deep reasoning mode errors reduced by 12.3%, speed mode errors reduced by 8.5%
  • Marvis Agent: Task completion rate reached 93.7%, up 12.7% from preview version; with 6 Agents collaborating, task scheduling accuracy reached 92%, up 13.5%

Benchmark test scores comprehensively validate these real-world performances. In Agent and tool orchestration capabilities, Hy3 scored 68.5 on ClawEval pass³, surpassing DeepSeek V4 Pro's 62.4 and Qwen 3.7 Max's 65.5; SkillsBench jumped from the preview version's 29.1 to 55.3; BrowseComp reached 84.2, approaching GPT 5.5's 84.4.

Code agent performance also showed significant progress: SWE-bench Pro improved from the preview version's 46.0 to 57.9, NL2repo from 35.3 to 45.6. Mathematical reasoning capabilities saw even more dramatic improvement, with MathArena Apex leaping from the preview version's 12.8 to 38.7.

According to Tencent's official data, at least 90% of major scenario Agent tasks can be successfully executed, covering data processing, document generation, research analysis, information retrieval, web development, and life automation domains. Even more impressive, Hy3 supports complex Agent workflows of up to 495 steps, including document processing, data analysis, knowledge retrieval, and MCP tool chain orchestration.

Product Ecosystem: From Enterprise Office to Consumer Entertainment

Technical capabilities must ultimately demonstrate value through product deployment. Tencent has chosen a strategy of deep integration across all business lines.

In enterprise office scenarios, WorkBuddy and CodeBuddy have integrated Hy3 as the default recommended model, supporting code development and office automation. Tencent internally explicitly encourages business units to prioritize using Hy3, with strategic weight shifting toward self-developed models.

Consumer application deployment is even more extensive. Yuanbao's launched Agent functionality allows users to generate PPT, Word, Excel, PDF, HTML, and other format files through natural language conversations; Marvis enhanced file editing/generation, diagnostics, and multi-Agent collaboration capabilities; even in WeGame's "Path of Exile: Advent," the game assistant has integrated Hy3.

Developer ecosystem building is progressing in parallel. The API has gone live on Tencent Cloud TokenHub, third-party developer platforms worldwide are gradually connecting, and open-source products like OpenClaw, OpenCode, and KiloCode already support Hy3.

Pricing Strategy: Opening the Market with Cost-Effectiveness

In pricing, Tencent continues Hunyuan's tradition of high cost-effectiveness. API call prices are:

  • Input: $0.14/million tokens (1 RMB)
  • Output: $0.56/million tokens (4 RMB)
  • Cache hit input: $0.035/million tokens (0.25 RMB)

This pricing benchmarks against domestic competitors like DeepSeek V4 Pro and Qwen 3.7 Max, maintaining a clear cost advantage.

Even more aggressive is the free strategy: Yuanbao's Agent functionality is completely free and open to users, and the open-source version adopts the commercially friendly Apache 2.0 license. This "free + open source" combination punch is clearly designed to rapidly expand the user base and developer ecosystem.

Industry Impact and Future Outlook

The Hy3 launch has reshaped China's AI large model competitive landscape. In the Yuanbao Agent Benchmark, Hy3 has surpassed mainstream domestic models like GLM 5.1 in comprehensive office and life service scenarios, with performance approaching international first-tier models like Claude Sonnet 4.6.

This release also validates a technical approach: mid-size parameter scale (21B activated) combined with MoE architecture and reinforcement learning can find the optimal balance between performance and efficiency. Going from infrastructure rebuild to official release in just six months demonstrates the rapid iteration capability of Chinese tech companies in the AI field.

Looking ahead, Agent capabilities are becoming the focal point of differentiated competition among large models, and Hy3's 90% task success rate has established a new industry benchmark. Through deep integration of its internal ecosystem—from office, social networking to gaming and customer service—Tencent is forming a unique application moat. The open-source plus free model may further accelerate developer ecosystem development.

From the self-critical "moving too slowly" to launching an Agent capability-leading flagship model six months later, Tencent Hunyuan has proven through action: in this AI marathon, recognizing problems and quickly adjusting is more important than always maintaining the lead.