Tencent said in its Hy4 repository that it is publishing weights, the trained numbers needed to run a model, for Hy4, an FP8 version and the base model. They are on Hugging Face, ModelScope and GitCode. TechNode reported the release came on Friday, after the November 4 launch. The model reads text, images and video.
Tencent's repository, the first-hand source, names the Tencent Hy License. TechNode reported Apache 2.0, which conflicts with it. Check Tencent's license file first. TechNode also cited a context window of about one million tokens. Tencent's repository text gives no such figure.
Hy4 has 770 billion parameters but uses 49 billion per token. It is a mixture-of-experts design: each layer has 256 specialist blocks, and a token uses 8 of them plus one shared block. Active parameters measure work per token, not memory needed, and the sources give no hardware figures. Open weights only matter once the license is read. Buyers should ask what the full model and the FP8 version need, and whether commercial use is allowed. The base model is not tuned for chat. The chat and reasoning use applies to Hy4 and Hy4-FP8. A built-in drafter keeps its cost constant on long inputs, Tencent says.