Tencent Hy

Tencent Hy
DeveloperTencent
ReleaseSeptember 2023; 3 years ago (2023-09)
TypeLarge language model
LicenseHy3 and Hy4 preview: Apache 2.0
Hy3 preview: Tencent Hy Community License
Websiteaistudio.tencent.com
Repositoryhttps://github.com/Tencent-Hunyuan

Tencent Hy (Chinese: 騰訊混元), formerly known as Tencent Hunyuan, is a family of large language models developed by the Chinese technology company Tencent.

History

Tencent introduced its Hunyuan foundation model in September 2023, initially offering it to enterprise customers through Tencent Cloud. The model became part of Tencent's broader push into generative artificial intelligence and was subsequently integrated into a growing number of the company's products and services. By October 2023, Hunyuan had been incorporated into more than 180 Tencent applications and services, including WeChat search, Tencent Meeting and Tencent Docs.

In March 2025, Tencent released Hunyuan T1, a reasoning language model based on the Transformer-Mamba architecture.

Tencent subsequently released a number of Hunyuan models with publicly available weights. In June 2025, it released Hunyuan-A13B, an 80-billion-parameter MoE model with 13 billion active parameters.

In 2026, Tencent began using the shortened Hy branding internationally, while retaining the Hunyuan name in the Chinese market. Hy3 preview was released in April 2026 as the first model produced using Tencent's rebuilt pre-training and reinforcement learning infrastructure.

Tencent released the completed Hy3 model on July 6, 2026. The company expanded access to the model internationally in August through products including WorkBuddy and Tencent Cloud TokenHub.

On August 28, Tencent released Hy4 preview, increasing the model size to 770 billion parameters and the context window to more than one million tokens. Tencent released Hy3 and Hy4 preview under the Apache License 2.0, while Hy3 preview was released under the Tencent Hy Community License.

Models

The Hy models are open-weight and use mixture of experts (MoE) architecturew, designed for tasks including software engineering, scientific research, financial analysis and the operation of AI agents.

The current Hy family includes Hy3 preview released in April 2026, Hy3 released in July 2026, and Hy4 preview released in August 2026. Hy4 preview has 770 billion total parameters, of which 49 billion are active for each token, and supports a context window of more than one million tokens.

Models
Model Released Total parameters Active parameters Context length
Hy3 preview April 2026 295 billion 21 billion 256K
Hy3 July 2026 295 billion 21 billion 256K
Hy4 preview August 2026 770 billion 49 billion >1 million

Hy3 preview

Hy3 preview was the precursor to Hy3. The completed Hy3 retained the same architecture and model size, but used expanded reinforcement learning and higher-quality post-training data to improve reasoning, agent capabilities, tool use and reliability.

Hy3

Hy3 is a mixture of experts model containing 295 billion parameters, with 21 billion parameters activated for each token. It supports a context length of 256,000 tokens and combines direct responses with a reasoning mode. Tencent developed the model for reasoning, long-context tasks, tool calling and agent-based applications.

The weights for Hy3 are distributed through Hugging Face, ModelScope and other model repositories under the Apache License 2.0.

Hy4 preview

Hy4 preview is a 770-billion-parameter MoE model with 49 billion active parameters. Its backbone contains 78 layers and 256 routed experts, with eight routed experts activated for each token in addition to a shared expert. It uses a sparse-attention architecture and supports a context length of approximately one million tokens.

Tencent designed Hy4 preview for multi-step tasks including software development, office work, financial analysis, game development and scientific research. The model supports tool calling and can be deployed using inference frameworks including vLLM and SGLang.

Usage and reception

Following its release, Hy4 preview became one of the most heavily used models on OpenRouter. For the seven-day period ending September 6, 2026, OpenRouter ranked Hy4 preview first among models on its platform by tokens processed, with 14.7 trillion tokens. Hy3 ranked seventh with 3.99 trillion tokens. OpenRouter states that its rankings measure usage on its own platform rather than model quality or usage across the overall AI market.

Reuters described the Hy4 release as part of Tencent's increased investment in artificial intelligence and its competition with other Chinese AI developers. Tencent has integrated models from the family into products including CodeBuddy, WorkBuddy, Yuanbao and other services.

See also

References

  1. ↑ "Hy3-preview". GitHub. Tencent. Retrieved September 8, 2026.
  2. ↑ "China's AI firms take spotlight with deals, low-cost models". Reuters. February 14, 2025. Retrieved September 8, 2026.
  3. ↑ "Tencent claims Hunyuan AI model surpasses GPT-3.5 in Chinese in recent upgrade". TechNode. October 27, 2023. Retrieved September 8, 2026.
  4. ↑ "Tencent upgrades AI model to compete with DeepSeek, Alibaba". www.msn.com. March 24, 2025. Retrieved March 31, 2025.
  5. ↑ "Hunyuan-A13B". GitHub. Tencent. Retrieved September 7, 2026.
  6. ↑ "Tencent Unveils Hy3 preview; Model Enhances Agent Capabilities and Real-World Usability". Tencent. April 24, 2026. Retrieved September 7, 2026.
  7. 1 2 3 4 "Tencent Hunyuan Officially Releases Hy3, Advancing Agent Capabilities and Deeper Product Integration". Tencent. July 6, 2026. Retrieved September 7, 2026.
  8. ↑ "Tencent Hy3 Now Available Globally, Extending Practical AI Across Products, Workflows and Cloud Services". Tencent. August 5, 2026. Retrieved September 7, 2026.
  9. 1 2 3 "China's Tencent releases new open-source AI model for coding, research tasks". Reuters. August 28, 2026. Retrieved September 7, 2026.
  10. ↑ Tencent. "Hy3". GitHub. Retrieved September 8, 2026.
  11. ↑ "Hy4-preview". GitHub. Tencent. Retrieved September 8, 2026.
  12. ↑ "Hy3-preview". GitHub. Tencent. Retrieved September 8, 2026.
  13. ↑ "Tencent Q2 revenue climbs 11% on AI-driven ad gains, but profit falls short". Reuters. August 12, 2026. Retrieved September 7, 2026.
  14. ↑ "Hy3-preview". GitHub. Tencent. Retrieved September 7, 2026.
  15. ↑ "Hy3". GitHub. Tencent. Retrieved September 7, 2026.
  16. ↑ "Hy4-preview". GitHub. Tencent. Retrieved September 7, 2026.
  17. ↑ "Hy3". Hugging Face. Tencent. Retrieved September 7, 2026.
  18. ↑ "Hy4-preview". Hugging Face. Tencent. Retrieved September 8, 2026.
  19. ↑ "AI Model Rankings". OpenRouter. OpenRouter, Inc. Retrieved September 7, 2026.