Tencent releases Hy4 frontier model weights – language backbone has 770 billion parameters
The instruction-tuned, FP8 and base versions of Hy4 are now available to download. Pandaily reported that Tencent’s own licence requires large operators offering the model to others as a general-purpose service to obtain separate permission.
Tencent released three versions of its Hy4 frontier model on 20 November. Hy4 is an instruction-tuned chat model, Hy4-FP8 is a version quantised to the lighter FP8 number format, and Hy4 base is a pre-trained base model, not a chat model. Tencent has offered Hy4 as a service since 4 November, the technology outlet Pandaily reported.
Tencent’s Hugging Face model card says Hy4 understands text, images and video and has a context window of 1,048,576 tokens. Its language backbone has 770 billion parameters, but 49 billion of them are active for each token. The difference comes from the mixture-of-experts (MoE) architecture: in every MoE layer, each token is routed to eight of 256 experts plus one shared expert.
Even an MoE model usually has to keep all its weights in memory. The model card’s vLLM and SGLang examples split Hy4-FP8 across eight GPUs. Pandaily said the instructions do not show that the model would run on an ordinary laptop.
The weights use Tencent Hy License 1.0, not Apache 2.0. In Pandaily’s reading, the licence’s additional commercial terms apply to operators that offer others inference or fine-tuning of the model as a general-purpose remote service. Such an operator needs separate written permission if its total revenue exceeds $50 million in any 12-month period. The revenue does not have to come from Hy4. Permission is also required before the model is used in a service with more than 100 million monthly active users in any calendar month.
Pandaily said that end-user services with limited features, and services that merely forward requests to a model service run by someone else, do not count as model services. The limits therefore do not apply to every commercial user. The internal-use exception is narrow, however: it requires that the model, models derived from it and their outputs are not made available to third parties.
Tencent regards Hy4 as one of the strongest open-weight models. The company nonetheless acknowledges a clear gap to the best proprietary models in its internal benchmarks, particularly in research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. Pandaily downloaded a small part of a weights file from each of the three repositories but did not download the full models or run them.
The weights are available from Hugging Face, ModelScope and GitCode. The model card includes instructions and ready-made containers for vLLM and SGLang, and a quantised GGUF version is also available for testing.
Read more: Models