On February 10, Tencent Hunyuan made a significant announcement regarding the release of its 'ultra-compact' model, HY-1.8B-2Bit, specifically tailored for consumer hardware applications. Leveraging the pioneering industrial-grade 2Bit on-device quantization technique, this model employs 2Bit quantization to condense its equivalent parameter count to a mere 0.3B, resulting in a memory footprint of only 600MB—significantly smaller than many commonly used mobile applications. When compared to its original precision counterpart, the HY-1.8B-2Bit model achieves a sixfold reduction in parameter count while simultaneously boosting generation speed by 2-3 times. This makes it exceptionally easy to deploy on edge devices without any performance constraints.

The model incorporates a quantization-aware training approach, integrating data optimization, elastic stretching quantization, and innovative training strategies. These elements collectively enable it to deliver performance on par with 4-bit PTQ models across various metrics, including mathematics, coding, and science, thereby fulfilling its design objective of being 'compact yet potent.' Presently, the HY-1.8B-2Bit model has been optimized for computing platforms like Arm and can operate seamlessly on mobile devices that support Arm SME2 technology.