Article

Two days after Tencent released Hy4 flash, it open-sourced the models, releasing all three versions at once; the flagship Hy4 will have to wait.

Beating AI Flash News: Tencent Huan Yuan has officially open-sourced Hy4 flash, just two days after the model's release. This time, the weights for three versions were released simultaneously:

• Hy4 flash: The official model fine-tuned with instructions, which can be directly used for dialogue, reasoning, and agent tasks.

• Hy4 flash-FP8: An 8-bit quantized version, primarily used to reduce GPU memory usage during deployment.

• Hy4 flash Base: A pre-trained base model without instruction fine-tuning, making it easy for developers to continue training and customization.

Hy4 flash has a total of 224B parameters, with 12B activated per inference. Even after FP8 quantization, the deployment threshold remains high. Tencent has made the models available for download on GitHub, Hugging Face, ModelScope, and GitCode, and the community has also created GGUF quantized versions to facilitate running on various hardware environments.

The model uses the Tencent Hy License, which allows developers to use and modify the weights according to the agreement, but comes with licensing conditions. Large-scale model API service providers must obtain separate authorization if they meet specified revenue or user scale thresholds.

As for the flagship Hy4 with 770B parameters, Tencent stated that it is still accelerating the open-source process and has not yet announced a specific timeline.