What new capabilities did Hy4 add?
- Tencent released Hy4 on 4 November, improving instruction following, long-context understanding, long-duration tasks, and deep research capabilities.
- The new model can handle image details, complex charts, and visual math problems. It can also operate computers through graphical interfaces and the command line, broadening the range of real-world tasks it can handle.
How did it perform on real-world task tests?
- In a Tencent-led office collaboration evaluation, Hy4 outperformed participating domestic models in 9 of 12 WorkBuddy tests, and performed better in human-AI co-editing. WeCom can be used for project tracking, task and meeting reminders, deadline alerts, and priority ranking.
- The financial evaluation covered 579 tasks, scored by 213 experts, including 153 external experts. Hy4 scored 3.04 and Hy4; DeepSeek-V4.1-Flash-1017 scored 2.97, while Kimi K3.1 Pro and GLM-5.3 scored 2.68. Its internal-test hallucination rate fell from 15.76% for Hy4 Preview to 5.45%.
- Compared with Hy4 Preview, Hy4’s GameDev framework and GameLogic scores increased by 19% and 10%, respectively. In TenureTrail, an internal research test covering seven disciplines and 123 tasks, Hy4 outperformed the participating open-source models.
How do the model’s capabilities connect with enterprise applications?
- Morgan Stanley believes wider adoption of WorkBuddy across more industries and deeper integration with WeCom and Buddy One could strengthen the feedback loop between product usage and model improvement, accelerate iteration, and support enterprise adoption of AI.
- API list prices are unchanged: per million cached input, uncached input, and output tokens, the prices are RMB 0.3, 6, and 18, respectively. Tencent plans to open-source the model weights in the near term and launch Hy4 Flash, with faster inference and lower costs.
How will Tencent narrow the model gap next?
- Tencent plans to scale up training and co-design more frequently with its products to narrow the gap with closed-source models.
- Morgan Stanley rates Tencent Overweight, with a target price of HKD 550. Its sum-of-the-parts valuation assigns HKD 470 per share to core businesses and HKD 80 per share to associate investments.