Tencent releases Hy4 flash weights in instruction, FP8 and base versions

Tencent has now made Hy4 flash weights available in three variants, making its text-and-image model available for deployment outside the hosted API. The company’s official repository lists an instruction-tuned model, an FP8 version and a pretrained base model.
The model card describes approximately 220 billion total parameters, including its large n-gram embedding component, with roughly 12 billion active per token. Hy4 flash accepts text and images, generates text and supports up to 1,048,576 tokens of context.
Tencent links downloads on Hugging Face, ModelScope and GitCode. Its documentation distinguishes the base checkpoint from the instruction-tuned versions: the base is not presented as a ready-made chat model. The repository also provides deployment and fine-tuning guidance.
The weights use the Tencent Hy License, rather than an unrestricted public-domain release. Developers should review those terms before deployment.
Availability was checked on November 18; the exact release date was not established. This updates our earlier API coverage, when weights were still planned. Downloadable weights give teams a way to evaluate and adapt the model in their own environments, although its size means deployment still requires substantial computing resources.
Sources and citations
Revision history
Published .