On November 17, 2026, Tencent Hunyuan released open weights for Hy4 flash in instruct, FP8 and base variants on Hugging Face, ModelScope and GitCode. The Hugging Face repository was created at 03:44 UTC on 2026-11-17.
Hy4 flash is a cost-oriented multimodal MoE model: about 220B total parameters (including 63.5B Engram, 0.43B ViT and a 0.85B DFly2 drafter), with roughly 11.3B text parameters plus 0.43B vision parameters active per token. Its 42-layer language backbone has one dense layer and 41 MoE layers, uses hybrid attention, LatentMoE and N-gram embeddings, and supports a context window of 1,048,576 tokens.
Tencent positions it for everyday productivity: deep search, document processing, slide creation, simple frontend demos and computer use; long-horizon software engineering is still better suited to Hy4. The weights use the Tencent Hy License, so check commercial and redistribution terms before use. Open weights are not a subscription quota reset; any effect on Tencent product quotas would be a separate announcement.