Tencent Hunyuan has announced the open-sourcing of its Hy4flash series models, comprising three versions: the standard Hy4flash, the quantized Hy4flash-FP8, and Hy4flash Base for researchers to train from scratch. These models are available for download on multiple platforms, with the GGUF format provided to facilitate local deployment. The official announcement also teased that the larger Hy4 model will be open-sourced soon. This series follows a cost-effectiveness approach, featuring a total parameter count of 224B—including 64B Engram parameters—while activating only 12B per token. It employs a sparse architecture that balances computing power savings with model capability; it supports multimodal input, offers a context length of up to 1 million tokens, and its API pricing is significantly lower than that of Hy4, making it the most cost-effective Hy model to date. Currently, multiple Tencent products have already integrated it. Its open-sourcing further lowers the threshold for large model application costs and can also leverage community feedback to benefit its own ecosystem.
