The image generated by AIYXL using AI
Tencent released the official version of Hunyuan Hy4 on November 4, and Tencent Cloud's Agent Development Platform ADP has integrated it. The new version adds visual understanding and computer operation capabilities; the official weights and system card remain to be disclosed.
When a chart is handed to a model to explain, what the user gets is a passage of text. When the model is asked to open a file, find the chart, modify it, and save it again, what the user gets is a modified file. Tencent released the official version of Hunyuan Hy4 on November 4, bringing the latter type of task within the capabilities of this flagship model. Tencent Cloud said that the Agent Development Platform ADP has integrated the official version, and developers can select it in the workspace.
Compared with the previous Hy4 Preview, Tencent said the official version expands the training scale, adds visual understanding for the first time, and improves instruction following, long-document understanding, and long-horizon task capabilities. Computer operation can combine the command line and graphical interface. For work that spans software, this means the model can first read screen information and then, depending on the environment, choose to proceed via commands or the interface. The tools actually available, the files accessible, and the actions that can be executed are still determined by the integration environment.
The model service context disclosed by Tencent Cloud is about 1.05 million tokens, with a maximum output of 65,536 tokens per request; the specific upper limits are subject to the access point. These two figures describe, respectively, the context the model can receive and the output it can generate. How long a workflow can run continuously also depends on tool responses, whether task state can be recovered, and whether intermediate operations require user confirmation. The capabilities and evaluations announced by Tencent are vendor disclosures, and performance in actual business scenarios still needs to be tested with one's own tasks.
For example, to aggregate charts from a batch of monthly reports into a presentation, you can first test with read-only materials and copies. During acceptance, check item by item whether the data can be traced back to the original tables, whether page numbers and references are updated accordingly, and whether rerunning generates duplicate files. Evaluating Hy4 with such tasks can show exactly which manual step the visual capability reduces and which steps still require human confirmation.
What can be confirmed at present is the release of the official version and the ADP integration. Tencent said it will make the official version weights available and release the system card in the near future. For teams planning self-deployment, these two materials will determine whether model behavior can be reproduced and under what license and technical conditions deployment should take place.