Researchers at Microsoft have announced the development of the largest single-bit model of artificial intelligence, an architectural approach called “bitnet.” The BitNet b1.58 2B4T model is open source under the MIT license and requires only a central processor, including the Apple M2, to run.
Bitnet systems are compressed models designed to run on low-end hardware. In the case of standard models, the weights—the values that define the model’s internal structure—are often quantized. Quantization reduces the number of bits needed to represent the weights, allowing the models to run faster on systems with less memory. Bitnet models quantize weights into three values: -1, 0, and 1, meaning that in theory they are much more memory- and computationally efficient than most modern AI systems.
BitNet b1.58 2B4T, according to Microsoft, is the first model based on this architecture to have 2 billion parameters, with parameters largely the same as weights. It was trained on a dataset of 4 trillion tokens, which is estimated to be equivalent to about 33 million books. BitNet b1.58 2B4T is on par with similarly sized models: it outperformed Meta
Llama 3.2 1B, Google Gemma 3 1B, and Alibaba Qwen 2.5 1.5B on GSM8K (primary school math) and PIQA (common sense assessment). At the same time, the model is twice as fast as its peers in some cases and uses less memory.
But there is one caveat: to achieve maximum performance of the model, the bitnet.cpp framework developed by Microsoft is required, which only supports certain hardware. The list of supported chips does not include graphics processors, without which the modern AI industry is unthinkable. Thus, the architectural approach of “bitnet” seems to be a promising direction, but the obstacle is hardware compatibility.