Tencent has released new open-source translation models called HY-MT 1.5. The models are available in two versions: 1.8B and 7B. Both versions can operate on devices and in the cloud, offering high speed and accuracy.
The models come in two sizes: 1) 8B, optimized for consumer hardware, requiring approximately 1 GB of memory. Using on-policy distillation, it achieves a latency of 0.18 seconds for 50 tokens, surpassing many commercial APIs. 2) 7B, a new state-of-the-art model, outperforms most medium-sized open-source models and is comparable to proprietary solutions. It supports 33 languages and 5 Chinese dialects with high accuracy. These models are already in use within Tencent’s services.
The combination of these two models ensures stable operation both on devices and in the cloud.
Try them out at: GitHub: https://github.com/Tencent-Hunyuan/HY-MT, Hugging Face: https://huggingface.co/collections/tencent/hy-mt15, and more at: https://hunyuan.tencent.com/modelSquare/home/list