Microsoft researchers have introduced BitNet b1.58 2B4T, the largest 1-bit quantized AI model developed to date, marking a milestone in memory-efficient model design. Released under the MIT license, the model is optimized for CPUs — including Apple’s M2 — without relying on GPUs, which typically dominate AI infrastructure.
BitNet models, or “bitnets,” compress neural network weights into just three values: -1, 0, and 1. This enables higher speed and reduced memory use, ideal for lightweight hardware. With 2 billion parameters trained on 4 trillion tokens, BitNet b1.58 2B4T reportedly outperforms similarly sized models from Meta, Google, and Alibaba on reasoning benchmarks like GSM8K and PIQA.
The model’s performance is unlocked using Microsoft’s custom framework, bitnet.cpp, currently limited to select hardware platforms.
