Microsoft researchers claim they have developed the largest 1-bit artificial intelligence model to date, also known asBitnets”. The model is called BitNet b1.58 2B4T and is licensed under the MIT licensePublicly available, can run on CPUs including Apple M2.
Bitnets are essentially a compression model designed to run on lightweight hardware. In standard models, weights (values that define the internal structure of the model) are often quantized so that the model performs well on a variety of machines. Quantizing weights reduces the number of bits (the smallest units a computer can process) needed to represent those weights, allowing models to run faster on chips with less memory.
Bitnets quantize weights into three values: -1, 0, and 1. In theory, this makes them more memory and computationally efficient than most models today.
Microsoft researchers say BitNet b1.58 2B4T is the first BitNet to have 2 billion parameters, with "parameters" being largely synonymous with "weights." The researchers claim that BitNet b1.58 2B4T, trained on a dataset of 4 trillion tokens (estimated to be equivalent to about 33 million books), outperforms traditional models of similar size.
To be clear, BitNet b1.58 2B4T doesn't exactly beat the competition's 2 billion parameter model, but it does seem to hold its own. According to the researchers' tests, the model outperformed Meta's Llama 3.2 1B, Google's Gemma 3 1B and Alibaba's Qwen 2.5 1.5B on benchmarks including GSM8K (a set of elementary school math questions) and PIQA (a test of physics common sense reasoning abilities).
Perhaps even more impressively, the BitNet b1.58 2B4T is faster than other similarly sized models—in some cases twice as fast—while using only a fraction of the memory.
However, there is a problem. Achieving this kind of performance requires the use of Microsoft's custom framework bitnet.cpp, which currently only works on certain hardware. GPUs, which dominate the world of AI infrastructure, have not yet been included in the list of supported chips.
In summary,BitnetsThe future may be bright, especially for resource-constrained devices, but compatibility is still a key issue and likely will remain so.