Tencent Officially Releases Hunyuan Hy3 with Free Agent Features and Open Source Access

Tencent announced on July 6, 2026, the release and open-sourcing of the official version of its large language model, “Hunyuan” Hy3. On the same day, the model was also integrated into Tencent’s messenger app, “Yuanbao,” with all Agent features available for free.

Hy3 employs the Mixture of Experts (MoE) architecture, combining fast and slow thinking processes. It boasts a total of 295 billion parameters, with 21 billion active parameters, and supports a maximum context length of 256K tokens. The model is optimized for both coding and agent scenarios.

Compared to the preview version released on April 23 this year, the official version features a 20–30% improvement in agent and coding capabilities. Tencent also reported that known issues have been resolved, resulting in significantly improved model stability and user experience.

Optimization by Use Case

Hy3 offers enhanced capabilities in two key Agent functionalities.

Enhanced Coding Agent Capabilities

Deep optimizations have been made for front-end development scenarios. Practical improvements have been realized in areas such as understanding front-end requirements, generating components, describing interaction logic, and performing automatic debugging. These enhancements are expected to boost productivity in front-end development.

Strengthened Working Agent

The model now excels in planning tasks for complex workflows, processing lengthy information, and invoking multiple tools. Its autonomous execution capabilities have become more reliable, and the accuracy of multi-stage calls has improved. It is now equipped to handle general business workflows.

Benchmark Performance

Hy3 has achieved high scores across multiple benchmarks. In the SWE-bench Pro test, which evaluates programming and engineering capabilities, it scored 57.9 points, surpassing DeepSeek V4 Pro and Seed2.1 Pro. In the FrontierScience-Olympiad, which includes complex reasoning and mathematical problems, Hy3 outperformed GPT 5.5 and Qwen3.7 Max, ranking first among all tested models.

However, Hy3 ranked fourth in the SWE-bench Multilingual programming test, Terminal Bench 2.1 for terminal operations, and BrowseComp for web interaction. In the AA-LCR long-context reasoning test for agents, it tied for second place with Qwen3.7 Max, behind GPT 5.5.

Tencent also conducted internal testing. In blind tests involving 270 experts performing real-world tasks, Hy3 achieved an average score of 2.67/4, surpassing GLM5.1, which averaged 2.51/4. Hy3 demonstrated significant advantages in categories such as front-end, data storage, and CI/CD.

Pricing and Integration

The API pricing for Hy3 is set at 1 yuan (approximately $0.14) per 1 million tokens for input, 4 yuan (approximately $0.57) per 1 million tokens for output, and 0.25 yuan per 1 million tokens for cache hits.

The Hy3 official version has been immediately integrated into Yuanbao, offering Agent features for free. Users can activate the Agent functionality by switching to Hunyuan’s fast-thinking mode, explaining their requirements during a conversation, and specifying the type of file they want to generate.

Real-World Testing Results

The practical testing conducted by Zhidongxi has also been reported.

Front-End Animation Generation

The first test involved generating a self-contained HTML5 Canvas animation simulating a physical phenomenon. The instruction was to create a scene of a train derailing from a broken bridge and falling into water.

Hy3 successfully understood the elements like “bridge collapse,” “train falling,” and “water surface” to generate a functioning Canvas animation. However, its physical simulation was deemed inconsistent. For instance, the depiction of the bridge’s collapse and the hierarchy of its breakdown were weak, and the relationship between train cars appeared unnatural. In the segment where the train fell into the water, some cars seemed to float on the surface, revealing issues with gravity and sinking dynamics.

For comparison, the results of DeepSeek-V4-Pro were also reviewed. The latter demonstrated better continuity, capturing details such as the train’s approach, the bridge’s collapse, and the sequential derailment and fall of the cars, along with splashes and smoke effects.

In a second test, the task was to generate a scene where two cars launched off a slope and collided mid-air over a canyon. The resulting animation was notably more polished, accurately depicting the slope, canyon, car launches, and mid-air collision.

Project Generation and Document Analysis

In a Working Agent test, Hy3 was tasked with planning a project and generating core code for a WeChat mini-program for food delivery. The output demonstrated high completeness, including a detailed README file. Backend APIs covered flows for store listings, store details, order creation, order details, order lists, mock payments, address CRUD, and order cancellations.

For a deep analysis report on SpaceX’s IPO filing, Hy3 generated a comprehensive report covering key information. While Yuanbao could generate a full PowerPoint overview, its visual quality required manual adjustments.

Editorial Opinion

The most noteworthy aspect of the Hy3 official release is its emphasis on advancing Agent capabilities. Tencent has solidified the foundational Agent features in the preview version and elevated them to a practical level in this official release. High scores on the SWE-bench Pro and superior performance in internal blind tests indicate Hy3’s growing applicability in real-world operations. In the short term, the free integration of Hy3 into Yuanbao is likely to accelerate the adoption of its Agent capabilities. In the competitive Chinese market, where DeepSeek and Qwen are also enhancing their Agent functionalities, Tencent’s free strategy appears aimed at gaining a competitive edge.

In the long run, the key challenge for Tencent Hunyuan will be to ensure Hy3’s reliability in handling longer task chains and completing complex operations stably. The physical simulation tests revealed areas for improvement, indicating that Hy3 is not yet ready as a fully autonomous agent. However, its open-source nature holds promise for community-driven improvements. As an editorial team, we believe the ultimate measure of Hy3 will be its reliability and versatility as an enterprise-level agent system in practical applications.

References

• Huxiu Original Article — Published July 6, 2026