On November 20, 2026, Tencent's Hy team open-sourced Hy4, Hy4-FP8 and the pretrained Hy4-base weights on Hugging Face. Hy4 is a new-generation Mixture-of-Experts flagship with a 770B-parameter language backbone, 49B text parameters and 1.23B vision parameters active per token, text/image/video understanding, and a 1,048,576-token context.
The 78-layer language backbone uses one dense FFN layer followed by 77 MoE layers, each with 256 routed experts and one shared expert; every token activates the top-8 routed experts. Attention uses Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache, residual connections use iHC (identity Hyper-Connections), and the model includes a 1.23B-parameter ViT plus a 3.3B-parameter DFly2 speculative-decoding drafter.
The repositories provide instruct, FP8 and base weights, covering text, image and video inputs as well as computer use, frontend and game development, finance and office workflows, scientific research and long-horizon personal assistants. ModelScope, GitCode and third-party GGUF entry points are listed, and the weights use the Tencent Hy License (tencent-hy-license).
Hy4 is an open-weights release, not a subscription quota reset; any effect on Tencent product quotas would be a separate announcement.