Tencent unveiled Hy4, adding image understanding and computer use for the first time compared with the Hy4 preview. Developers and business users can assign image and screen tasks as well as long tasks spanning multiple apps. It is now available through Tencent Cloud, OpenRouter and other channels.
1. Connecting information read from images to screen operations
Hy4 scales up pretraining and post-training, the stages of learning and refinement, compared with the earlier preview. It adds vision and computer use for the first time. Tencent says it can read image details and complex charts and operate both command lines, where commands are typed as text, and graphical interfaces.
Screen-operation training included professional software such as Blender and CAD, tools for creating designs on computers. Training also covered workflows across applications to enable task planning and execution.
Tencent reports OSWorld 2.1 computer-use scores of 78.9 for Hy4 and 79.6 for Kimi K3.1 Pro under the Full, partial score condition. Each model was evaluated at its highest available reasoning setting.
2. Training on game engines and multiplayer environments
For game development, Tencent worked with developers from its Interactive Entertainment Group to create training materials and evaluations. It adapted the model to major game engines, the foundations running game rules and visuals, and covered multiplayer game development. The company reports gains over the preview of 19.1% in game development evaluation and 9.8% in game logic evaluation.
3. Performing research tasks across multiple stages
Long-task capability focuses on sustained planning and execution rather than answering a single question. Tencent presented TenureTrail, an internal evaluation of 123 executable research tasks across 7 fields, including mathematics, physical sciences and life sciences. The company says Hy4 ranked first among models it classified as open-weight, meaning models that disclose their internal numerical parameters.
Table: Research task evaluation scores
Conditions: Internal TenureTrail evaluation; 123 executable research tasks across 7 fields; each model's highest reasoning setting, governing how it computes answers; Claude Code 2.1.220; as reported by Tencent
Unit: points
| Model | Score |
|---|---|
| Hy4 | 53.6 |
| Kimi K3.1 Pro | 51.3 |
| DeepSeek-V4.1-Flash-1017 | 50.1 |
Hy4 led the three models evaluated with the same execution tool. Tencent bases its long-task claims on such research evaluations and training in workflows spanning multiple apps.
4. Improving search and document quality for finance and office work
The finance evaluation covered 579 real-world financial tasks involving 213 experts. The company says Hy4 ranked first among models it classified as open-weight. Reported scores for the same tasks are shown below.
Table: Real-world financial task evaluation scores
Conditions: 579 real-world financial tasks involving 213 financial experts; as reported by Tencent
Unit: points
| Model | Score |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
Hy4 scored slightly higher than its closest comparison model in the finance evaluation.
The company says it improved search to retrieve up-to-date information, find and fetch recent materials, synthesize multiple sources and check factual consistency. In its internal evaluation, the rate of fabricated content fell from the preview's 15.76% to 5.45%, while speculation errors fell from 23.1% to 5.6%.
For document work, Manus reported that Hy4 completed about 97% of generation and editing tasks in an internal evaluation.
Investment research firm RabyteTech said Hy4 scored 12.03 percentage points higher than the preview across 101 tasks in its internal evaluation.
5. Initial access through APIs and workplace products
Developers can access Hy4 through Tencent Cloud or OpenRouter APIs. Business users can access it through products incorporating it, such as WorkBuddy and Yuanbao.
WorkBuddy connects to enterprise messenger WeCom to organize project updates and to-dos from chats, emails, documents, schedules and meeting records. Long-task use therefore extends beyond research to everyday project management.
The company says API pricing is unchanged from the preview. Users gain image understanding and screen operations at the same rates, but must pay API fees or connect their work through products that incorporate those features.
With model weights planned for later release, deployment on users' own infrastructure, servers and equipment remains a next step. Tencent also plans to publish a system card at release, documenting safety evaluations, technical details, capabilities and safety information. The company acknowledges remaining gaps with leading closed models in areas including long-horizon coding and advanced visual question answering.
Hy4 moves beyond conversation to task execution combining visual information and multiple tools, with access currently available through Tencent products and APIs.