Tencent releases Hy4 open weights: 770B-parameter flagship with new licence terms

Tencent on November 20 released the full weights of its flagship model Hy4 on Hugging Face, ModelScope and GitCode, including the instruction-tuned version, an FP8 quantised version and the pre-trained base, along with a system card. The official weights use the new Tencent Hy License, under which large model API providers need separate written authorisation.
Tencent has released the full weights of its flagship large model Hy4 for download, available on Hugging Face, ModelScope and GitCode. The release includes the instruction-tuned Hy4, the FP8-quantised Hy4-FP8 and the pre-trained base model Hy4-base, along with a system card describing the model's capabilities and limits. It was announced at 03:52 UTC on November 20, or 22:52 EST on November 19.
770 billion parameters, a million-token window
Hy4 uses a mixture-of-experts (MoE) architecture with 770 billion total parameters, about 49 billion activated per token, 78 layers and a context window of about one million tokens. It supports text, image and video understanding as well as computer use, and Tencent has provided deployment guides for vLLM and SGLang so developers can run it on their own hardware. The model itself is unchanged from its official launch as an API service on November 4.
New licence: large providers need separate authorisation
Notably, the licence for the official weights differs from the August preview, which used Apache 2.0. The full release uses the new Tencent Hy License. Derived from Apache 2.0, it adds a condition for large model API providers: products with annual revenue above US$50 million or more than 100 million monthly active users need separate written authorisation from Tencent. Some reports still describe the release as Apache 2.0, which conflicts with the model card; the model card governs.
From API to weights in a month
Hy4 was first released as an API on November 4, with Tencent promising the weights would follow; on November 18 the lighter Hy4 Flash, with 220 billion parameters, was open-sourced first. With the flagship weights now open, developers no longer need to use Hy4 only through Tencent's API and can deploy, fine-tune and study it locally. The release adds to the intensifying competition between open-weight frontier models and closed API services.
N環球 
💬 Comments
No comments yet. Be the first!