Lates News

Based on the first industry-level 2Bit edge quantization solution, Tencent Hybrids officially launched a "ultra-small" model HY-1.8B-2Bit today, targeting consumer-level hardware scenarios. The equivalent parameters of the model are only 0.3B memory usage, with only 600MB, which is even smaller than some commonly used mobile applications. By using 2-bit quantization aware training (QAT) on the previously small-sized language model HY-1.8B-Instruct, this model has reduced equivalent parameters by 6 times compared to the original precision model, while also maintaining the full thinking ability of the original model. In addition, the speed of generating the model on real edge devices has been increased by 2-3 times compared to the original precision model, significantly improving the user experience. Tencent Hybrids' release of HY-1.8B-2Bit model allows stress-free deployment on edge devices, making it the first practical implementation of 2-bit industry-level quantization on edge models.