Microsoft has made waves in the AI community by unveiling BitNet b1.58 2B4T, the largest-scale 1-bit AI model to date. This model, designed to run on lightweight hardware such as CPUs, is now available for use under an MIT license. While Bitnets are not new, BitNet b1.58 2B4T pushes the boundaries of what’s possible by offering incredible memory and computational efficiency while outperforming other models of similar sizes in key benchmarks.
Bitnets are a type of compressed AI model that reduces the memory requirements of traditional models. In standard AI models, the weights — or parameters — that define the internal structure of the model are quantized, meaning they are reduced to a smaller set of values to make the model more efficient. However, traditional quantization usually involves multiple values. A BitNet takes this process even further, using only three possible values: -1, 0, and 1, significantly lowering the memory and computational requirements.
In theory, this allows Bitnets to run on lightweight hardware with fewer resources, making them ideal for resource-constrained devices like smartphones and embedded systems.
The newly developed BitNet b1.58 2B4T is the first BitNet to feature 2 billion parameters. This makes it one of the largest-scale Bitnets to date. Trained on a dataset of 4 trillion tokens (roughly the equivalent of 33 million books), the model demonstrates impressive performance and speed despite its compressed nature.
According to Microsoft’s researchers, BitNet b1.58 2B4T manages to outperform comparable models in benchmark tests like GSM8K (which involves grade-school-level math problems) and PIQA (which tests physical commonsense reasoning). Specifically, it outperforms Meta’s Llama 3.2 1B, Google’s Gemma 3 1B, and Alibaba’s Qwen 2.5 1.5B on these tasks.
Moreover, the model is twice as fast as other similar models, all while using a fraction of the memory typically required.
However, there’s a catch to all of this impressive performance. To run BitNet b1.58 2B4T, you need to use Microsoft’s custom framework, bitnet.cpp, which currently only supports certain hardware, most notably CPUs like Apple’s M2 chip. The model does not support GPUs, which are the dominant hardware used in modern AI infrastructure. This limitation means that while the model holds promise for lightweight devices, it may not be practical for large-scale deployment on the most widely used AI hardware.
