Microsoft researchers have introduced BitNet b1.58 2B4T, the largest 1-bit AI model to date, available under an MIT license. Designed for CPU compatibility, including Apple’s M2, this model utilizes a compressed format that quantizes weights into three distinct values, enhancing efficiency. With 2 billion parameters trained on a dataset equivalent to 33 million books, it outperforms several traditional models. While it excels in speed and memory use, compatibility limitations with existing GPU infrastructure remain a concern, potentially hindering widespread adoption.