Microsoft researchers say they have created a new kind of AI model that is very efficient and can run on regular computer processors (CPUs), like Apple’s M2 chip. The model is called BitNet b1.58 2B4T and it is available for free under an MIT license.
This AI model is a type of “bitnet”, which is a smaller, compressed version of AI designed to work on devices with less memory and power. Most AI models use a lot of memory to store their “weights” (the values that help the AI learn), but bitnets reduce these weights to just -1, 0, or 1. This helps the model run faster and use less memory.
BitNet b1.58 2B4T is the largest bitnet so far, with 2 billion parameters (which are similar to weights). It was trained on a huge amount of data — equal to around 33 million books.
In tests, this bitnet did better than other similar AI models from companies like Meta, Google, and Alibaba on tasks like math and logical reasoning. It also runs faster — sometimes twice as fast — and uses much less memory.
However, there is one limitation. To get the best performance, the AI model needs to use Microsoft’s special software called bitnet.cpp, which only works with certain types of hardware. Right now, it doesn’t work with GPUs, the powerful chips most AI systems use.
Still, this new model shows a lot of promise, especially for devices with limited resources, like laptops or small servers. But hardware support will be an important issue moving forward.
Source: techcrunch
