[CLSTR]
< Back to situation

[REVISION HISTORY]

Tencent Hy4 preview AI model release and deployment

Updated 2 times since CLSTR started tracking revisions of this situation.

What changed

2026-09-03 11:06 UTC → 2026-09-09 09:38 UTC · added removed

Tencent has released and open-sourced the Hy4 preview, a next-generation large language model (LLM) utilizing a Mixture-of-Experts (MoE) architecture. The model features 770 billion total parameters, with 49 billion parameters activated per token, and supports a context length of over 1 million tokens. Released under the Apache License 2.0, the model is designed Designed to enhance productivity in specialized fields such as software engineering, scientific research, and financial analysis. office tasks, the model was co-designed with Tencent products such as WorkBuddy and CodeBuddy. In an internal blind evaluation involving 163 experts and 203 engineering tasks, the Hy4 preview achieved an average score of 2.99 out of 4.00, outperforming models including GLM-5.3 and Kimi K3. Following its integration into the WorkBuddy AI service on August 28, Tencent previously announced an emergency expansion of its inference clusters to address a surge in demand. High high request volumes have resulted in task queues and waiting periods. task queues. To manage the load, the company encourage adoption, Tencent is dynamically allocating resources and has suggested users utilize the Hy3 model or access services during off-peak hours to avoid delays. A offering a free trial period for the Hy4 preview through WorkBuddy and CodeBuddy. The model is scheduled to run until September 10, with daily usage limits applied per user. accessible globally via Tencent applications including WorkBuddy, CodeBuddy, Yuanbao, and ima. Developers can also access the model via APIs through Tencent Cloud TokenHub and OpenRouter. In recent benchmarks, the Hy4 preview demonstrated significant improvements over its predecessor, Hy3. On Hy3, ranking eighth globally on the Code Arena WebDev leaderboard, the model ranked eighth globally, placing just behind Anthropic’s Claude Fable 5 and outperforming Alibaba Group Holding’s Qwen 3.8-Flash-Next. Analysts from Goldman Sachs noted that Tencent’s approach involves a “differentiated product-plus-model strategy,” using its product ecosystem to collect user data for training cycles. The model is accessible via WorkBuddy, CodeBuddy, and through APIs via Tencent Cloud TokenHub and OpenRouter. leaderboard.

Versions

  1. 2026-09-09 09:38 UTC Tencent Hy4 preview AI model release and deployment
  2. 2026-09-03 11:06 UTC Tencent Hy4 preview AI model release and deployment
  3. 2026-08-31 05:09 UTC Tencent Hy4 preview AI model release and deployment

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.