A Guth Labs publication ↗
Back to News

Tencent releases the weights for its flagship Hy4 model

Three interlocking cream ceramic and green glass blocks with a thin gold strand tracing their seams.
AI-generated editorial illustration

Tencent released downloadable weights for its flagship Hy4 model on November 20, following its hosted launch earlier in the month. The model card lists an instruction-tuned model, a reduced-precision FP8 version of that model, and a pretrained base checkpoint. This is the full Hy4 model, not the smaller Hy4 Flash whose weights Tencent released separately.

Hy4 can interpret text, images and video. Tencent says the language backbone has 770 billion parameters, with 49 billion active per text token, and supports a context window of 1,048,576 tokens. Only part of the model handles each token, but a deployment still has to accommodate substantially more weights than the active count suggests.

Tencent links the checkpoints from Hugging Face, ModelScope and GitCode, and has published a system card. The weights use the Tencent Hy License, a customized license rather than unmodified Apache 2.0.

The release gives researchers and companies a way to evaluate the full Hy4 model outside Tencent’s hosted products, if they have the computing resources to run it.

Sources and citations

Revision history

Published .