Key takeaways
- Tencent published Hy4, Hy4-FP8 and Hy4-base weights on Hugging Face, ModelScope and GitCode.
- Hy4 is a mixture-of-experts model with 770 billion parameters, 49 billion active per token and a 1,048,576-token context.
- Unlike the Apache 2.0 preview, it uses the Tencent Hy License, which requires a separate license for model-as-a-service businesses above $50 million in revenue or 100 million monthly users.
Tencent has published the full weights of Hy4, its flagship model, on Hugging Face. Hy4 is a mixture-of-experts model, which runs only part of its network for each token: 770 billion parameters in total, 49 billion active per token, and a context window of 1,048,576 tokens. The chat model, an FP8 copy that needs less memory, and the pre-trained base model are all up, and the model pages were last updated on November 20.
This is the big sibling of Hy4 flash, the 220-billion-parameter model Tencent released earlier this week. It is also the finished version of the Hy4 preview that came out in August. One thing changed on the way: the license.
What Tencent released
The model card lists three downloads: Hy4, the instruction-tuned model; Hy4-FP8; and Hy4-base, the model before chat training. They are also on ModelScope and GitCode.
Hy4 reads text, images and video and writes text. The language model has 78 layers, each with 256 small expert networks, of which 8 plus one shared expert handle each token. A 1.23-billion-parameter vision encoder turns images and video frames into tokens. Like DeepSeek's recent models, it uses sparse attention, which looks at only the 2,048 most relevant earlier tokens instead of all of them. That is how it keeps a million-token context affordable.
A small built-in drafter, DFly2, with 3.3 billion parameters, guesses several tokens ahead so the big model can check them in one pass. Tencent ships it ready for speculative decoding in vLLM and SGLang.
How it compares, by Tencent's numbers
The scores published so far come from Tencent. On its base-model tests, Hy4 base scores 91.6 on MMLU and 73.9 on MMLU-Pro, ahead of DeepSeek-V4-Pro-Base at 91.1 and 73.4. On SimpleQA, a test of factual recall, it scores 57.2 against 52.1. DeepSeek's base model wins on the code test MBPP-plus, 79.1 to 74.0.
Tencent is open about the limits. The model card says "a notable gap remains" between Hy4 and the top proprietary models on research-level math, long-horizon coding and hard visual questions. No independent scores for the final Hy4 have been published yet.
Open weights, with a revenue limit
The Hy4 preview came under Apache 2.0, which lets anyone use it commercially with few conditions. The final Hy4 uses the Tencent Hy License instead.
Most users won't notice. The catch is for companies that sell model access as a service, such as an API. They need a separate license from Tencent if their worldwide revenue tops $50 million in any 12 months, or before using Hy4 in a product with more than 100 million monthly users. Using the model only inside a company, without sharing it or its outputs with others, is exempt.
That is a direct limit on cloud providers that want to host Hy4 for customers, which is how most people use large open models.
What you need to run it
This is a data-center model. Tencent's vLLM and SGLang examples serve the FP8 version across eight GPUs. A GGUF build from Tencent's AngelSlim team is available "for end-to-end testing purposes," the model card says. Tencent also publishes a fine-tuning pipeline and a system card.
The next thing to watch is independent testing. Artificial Analysis and OpenRouter usage will show whether Hy4 holds up against DeepSeek and GLM outside Tencent's own tables. Our page on Hy4 through the API and OpenRouter covers the hosted version.
Sources
- tencent/Hy4 model cardTencent
- Tencent Hy LicenseTencent
- tencent/Hy4-FP8Tencent
- tencent/Hy4-baseTencent
- Tencent Releases Hy4 Flagship Weights Under Conditional LicensePandaily