On November 18, 2026, Tencent's Hy team published Hy4-base on Hugging Face. It is the pretrained base version of the new-generation Mixture-of-Experts flagship model, supporting text, image and video understanding, with a language backbone totaling 770B parameters. Model-card tags include image-text-to-text, MoE, multimodal and pretrained.
The model card covers pretrained capabilities, key improvements, capabilities and use cases, known limitations and an appendix of benchmark results, plus quickstart, deployment, finetuning and System Card sections. Listed use cases include vision and computer use, frontend and game development, finance and office workflows, scientific research and AI R&D, and long-horizon personal assistants.
Weights are hosted on Hugging Face under the Tencent Hy License (tencent-hy-license), with corresponding entries on ModelScope and GitCode and a link to a third-party GGUF build. The deployment section provides vLLM and SGLang launch and call examples, alongside notes on the DFly2 speculative decoding component and quantization.
Hy4-base is an open-weights release, not a subscription quota reset; any effect on Tencent product quotas would be a separate announcement.