Microsoft researchers say they have developed a very efficient AI model that can work on central processing units

Microsoft researchers claim that they have developed the largest 1 -bit artificial intelligence model, also known as “bitnet”, so far. Bitnet B1.58 2B4T is called it Available publicly Under a MIT license and can be operated on CPU, including Apple M2.

Bitnets are mainly compressed models designed to run on lightweight devices. In standard models, weights are often and values ​​that determine the inner structure of the model The amount so that the models are well caused by a wide range of machines. Determining weights reduces the number of bits – the smallest units that a computer can process – needed to represent these weights, allowing models to operate on chips with less memory, faster.

BitNets performs the weights of weights to only three values: -1, 0 and 1. In theory, which makes them more efficient in memory and computing than most models today.

Microsoft researchers say Bitnet B1.58 2B4T is the first bitnet with 2 billion teachers, “parameters” as they are largely synonymous with “weights”. Training on a collection of data of 4 trillion symbols – equivalent to about 33 million books, According to one estimate Bitnet B1.58 2B4T surpasses traditional models of similar sizes, as researchers claim.

Bitnet B1.58 2B4T does not sweep the ground with 2 billion competing forms of the teacher, to be clear, but it seems to be pregnant. According to the researchers test, the Llama 3.2 1B model from Meta, Google Gemma 3 1B, and QWEN 2.5 1.5b from Alibaba exceeds the standards including GSM8K (a group of mathematics problems at the school level).

Perhaps more impressive, Bitnet B1.58 2B4T is faster than other models of their size – in some cases, poor speed – while using a part of the memory.

There is hunting, however.

Achieving this performance requires the use of the designated microsoft framework, bitnet.cPP, which only works with certain devices at the present time. Absolutely absent from the support list of graphics processing units, which dominate the Amnesty International’s infrastructure scene.

This means that bitnets may carry a promise, especially for the resource -bound devices. But compatibility – it is likely to remain – a large attached point.