ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Tencent Hunyuan Hy4 Open Weights: 770B Flagship Model Free to Download — With One Commercial Catch

Tencent Hunyuan Hy4 Open Weights: 770B Flagship Model Free to Download — With One Commercial Catch

AI information • Admin • • 10 views

On November 20, 2026, Tencent Hunyuan announced on its official account that it is releasing the open weights of its flagship model, Hy4. This fulfills a promise made when Hy4 launched on November 4: back then the model was API-only and the weights were nowhere to be seen; now the full version can finally be downloaded. For developers, that means running a 770-billion-parameter flagship model on their own machines instead of calling it through a cloud vendor.

What the 770B flagship release actually includes

Hy4 is a mixture-of-experts model with 770 billion total parameters, 49 billion activated per token, and a context window of 1,048,576 tokens, handling text, images, and video together. Tencent released more than one checkpoint: a standard BF16 version, an FP8 version aimed at production deployment, an un-instruction-tuned Base version, plus a community GGUF build, along with the system card carrying its safety evaluations. Deployment follows two official routes, vLLM and SGLang, and the FP8 version runs on eight GPUs with tensor parallelism set to eight. The model also carries two cost-saving designs: sparse attention plus an index cache to hold down long-context costs, and the DFly2 speculative decoder to push speed higher. Tencent's own tests say Hy4 Base leads DeepSeek-V4-Pro-Base on most knowledge and reasoning benchmarks — vendor-tested numbers that no third party has reproduced yet. The lightweight Hy4 Flash was already open-sourced two days earlier; this time it is the flagship itself.

Free to download does not mean free for any commercial use

The license is the part to read carefully. Hy4 uses Tencent's own Tencent Hy License, not a standard Apache 2.0: the weights can be downloaded for free, deployed locally, and fine-tuned, but anyone serving the model to others with annual revenue above US$50 million, or a product with more than 100 million monthly active users, needs a separate written license from Tencent. In other words, personal research and small-team self-hosting are essentially unrestricted, but large-scale commercial serving requires checking that line first. The system card also documents risk assessments for areas such as cybersecurity and biosecurity — worth reading before any commercial use.

Who can start using it right now

Teams with eight high-end GPUs can run the FP8 version today using the official vLLM or SGLang images; developers with only consumer-grade cards can try the GGUF build first. The weights are already listed on Hugging Face, ModelScope, and GitCode. Two things are worth watching next: whether the community produces serious fine-tuned versions built on Hy4, and whether independent benchmarks reproduce Tencent's results — those two will decide whether this is just another very large open model or one that is genuinely usable.

Recommended Tools

More