Tencent has released the open weights for Hy4 flash, its most cost-efficient model in the Hy family, and for the flagship Hy4. The repositories are available on Hugging Face, ModelScope and GitCode under the Tencent Hy License.
Hy4 flash: 220 billion parameters, 12 billion active
According to the official model card, Hy4 flash has around 220 billion total parameters (including approximately 63.5 billion in the Engram n-gram component), with about 12 billion active per token (11.3 billion text + 0.43 billion vision). It accepts text and image inputs, generates text, and supports a context of up to 1 million tokens.
The architecture includes hybrid attention (sparse + sliding window), LatentMoE and Engram. Tencent states it delivers performance close to or above Hy4 Preview on WorkIndex tasks at a much lower cost per task — roughly 1/13 of Preview and 1/20 of full Hy4.
Weights were published around November 17-18, 2026, including instruct, FP8 and base (pretrained) versions. Available via API on Tencent Cloud TokenHub and OpenRouter.
Hy4 flagship: 770 billion parameters
The Hy4 model, initially launched on November 4, had its weights opened around November 20. Reports and model cards list roughly 770 billion total parameters and 49 billion active, also multimodal with 1 million token context. It includes vision and computer-use capabilities.
Tencent positions Hy4 for complex long-horizon tasks such as software engineering, while recommending flash for everyday productivity (deep search, documents, slides, computer use).
Why it matters
The release of weights for large-scale Chinese MoE models expands open-weight options for developers, especially in cost-sensitive and long-context scenarios. Tencent’s custom license (not standard Apache/MIT) requires attention to commercial-use restrictions.
Sources: Hugging Face Hy4-flash, official Tencent announcement.
This article was produced with AI assistance and editorially reviewed.
Por Grok