Tencent Hunyuan announced today that it is open-sourcing the Hy4flash series, releasing three versions at once: the standard Hy4flash, the quantized Hy4flash-FP8, and Hy4flash Base for researchers to train from scratch. It can be downloaded from GitHub, Hugging Face, ModelScope, and GitCode, and a GGUF format is also provided to make it easier for local deployment users to run it. The official announcement also teased that a larger Hy4 model is on the way to being open-sourced.

The selling point isn't size—it's value
This model takes the cost-effectiveness route: it has 224B total parameters, including 64B Engram parameters, but activates only 12B per token. It uses a sparse architecture, saving compute while retaining capability. It supports multimodal input, has a context length of up to 1 million tokens, and its API price is significantly lower than Hy4—the official label calls it the most cost-effective Hy model to date.
Its adoption has also been broad: Tencent Cloud TokenHub, Yuanbao, WeCom, QQ, ima, Marvis, and other products have already integrated the model. For developers, the weights are downloadable, the license is open, and the price is lower—effectively lowering the cost barrier for large model applications yet again. For Tencent, open-sourcing is both a way to boost its visibility and a familiar play to use the community to feed back into its own ecosystem.
Join Now