Key points
- Hy4 is a 770-billion-parameter mixture-of-experts model with native vision and computer-use capabilities.
- Internal tests showed a 19.1% gain in game development benchmarks over the previous preview version.
- In financial services evaluations, Hy4 scored 3.04, outperforming DeepSeek-V4.1-Flash and Kimi K3.1 Pro.
The claim
Tencent states that Hy4 moves beyond simple response generation toward agentic software manipulation. Hy4’s architecture includes a 1-million-token context window, enabling the processing of large code repositories and intricate operational histories.Tencent’s new AI model can see screens and control computer programs, not just write text.Tencent’s 770B MoE model introduces native tool-use and vision capabilities for agentic workflows.
The response
Tencent stated that Hy4 achieved a 19.1% improvement on internal game development benchmarks compared to the preview version. Tencent tested Hy4 in the financial sector using 579 tasks and 213 industry experts. Hy4 achieved a score of 3.04, surpassing DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro at 2.68.
The bottom line for teams
Tencent acknowledged limitations in Hy4’s performance. The key issue for enterprises and programmers is whether these features are dependable enough for actual operational use.
What the paper shows
The research tests Hy4’s ability to operate software rather than just answer questions. It reports specific gains in game engine integration and financial task scoring against named open-weight rivals.reported The stated limitations include deficits in visual QA, complex math and extended software engineering compared to proprietary leaders. Uncertainty remains regarding reliability in untested production environments.
Why it matters
- Model evaluators can compare Hy4’s 3.04 financial score against DeepSeek-V4.1-Flash and Kimi K3.1 Pro for industry-specific benchmarking.
- The 19.1% internal gain in game development suggests potential utility for Unreal, Unity and Godot workflows.
- Acknowledged deficits in visual QA and long-horizon engineering help scope the model’s current operational boundaries.
- The one-million-token context window allows testing of large code repositories in agentic scenarios.
Updated
Then / now / next
Tencent offered a text-only preview of its AI model.
The full Hy4 model includes vision and computer-use capabilities for desktop applications.
Teams must determine if the model’s agentic skills are reliable for real-world production workflows.
Story timeline
- Then Tencent offered a text-only preview of its AI model.
- Now The full Hy4 model includes vision and computer-use capabilities for desktop applications.
- Next Teams must determine if the model’s agentic skills are reliable for real-world production workflows.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
How does Hy4 compare to other models in financial tasks?
Tencent reported Hy4 scored 3.04, leading DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro at 2.68 in evaluations with 213 professionals.
Which game engines does Hy4 support?
Tencent adapted Hy4 for Unreal Engine, Unity and Godot through collaboration with its Interactive Entertainment Group.
Reading receipt
- Hy4 is a 770-billion-parameter mixture-of-experts model with native vision and computer-use capabilities.
- Internal tests showed a 19.1% gain in game development benchmarks over the previous preview version.
- In financial services evaluations, Hy4 scored 3.04, outperforming DeepSeek-V4.1-Flash and Kimi K3.1 Pro.