The News, Explained
Tencent announced Hy4 on November 4, introducing image understanding. It plans to release model weights and system cards in the coming days. Weights are learned model data, while system cards describe model properties and evaluation information. Those public files remain at the planned-release stage. Source
Tencent Cloud TokenHub’s dated update log lists support for Hy4 on November 4 and Hy4 preview on August 28. These entries distinguish the current model announcement from the earlier preview, using each support record rather than the document’s update timestamp. Source
The same TokenHub directory gives both models identical context and input/output limits. A token is a unit of information processed by a model. Reading the context window alongside the separately listed input and output ceilings helps explain the service’s stated processing ranges. Source
| TokenHub specification | Hy4 preview | Hy4 |
|---|---|---|
| Context window | 1M tokens | 1M tokens |
| Maximum input | 960k tokens | 960k tokens |
| Maximum output | 64k tokens | 64k tokens |
The directory also lists image understanding among Hy4’s capabilities. This separates a new input capability from the size of the listed processing ranges. Source
OYOPICK’s Take
The interesting change is the kind of material the model can handle while its listed ranges stay the same. Pairing a picture with a question and checking whether the explanation correctly connects its labels and relationships would turn the image-understanding claim into a concrete task to evaluate.
Once weights and system cards are published, we hope to see outside evaluations using the same inputs and clearly stated test conditions. Such evidence would make it easier to judge how much image understanding helps people interpret material.