Tencent open-sourced Hy4 flash two days after its release, rolling out three versions at once, while the flagship Hy4 is still awaited.
Beating AI News Flash: Tencent Hunyuan has officially open-sourced Hy4 flash, just two days after the model's release. This time, three versions of model weights have been released at once:
• Hy4 flash: The official model fine-tuned with instructions, which can be directly used for dialogue, reasoning, and agent tasks.
• Hy4 flash-FP8: The 8-bit quantized version, mainly used to reduce GPU memory usage during model deployment.
• Hy4 flash Base: The pre-trained base model without instruction fine-tuning, making it convenient 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 quite high. Tencent has opened downloads on GitHub, Hugging Face, ModelScope, and GitCode, and the community has also created GGUF quantized versions for running in different hardware environments.
The model uses the Tencent Hy License, allowing developers to use and modify the weights according to the agreement, but with licensing conditions attached. Large model API service providers that reach the specified revenue or user scale thresholds need to obtain separate authorization.
As for the flagship Hy4 with 770B parameters, Tencent says it is still accelerating open-sourcing and has not announced a specific timeline.