Microsoft researchers have claimed that they have created the largest scale 1-bit AI model, it is also known as “bitnet” to today. BitNet B 1.58 2B4 is called, it is Publicly available Can run under a MIT license and on CPUs with Apple’s M2.
Bitnets are essentially compressed models designed to run light weight hardware. In standard models, in weight, values that define the internal structure of a model, are often Quantized so models perform well on a wide machine. Reduces the number of bites that have determined the amount of weight – the smallest units of a computer can process – those weights are needed to present, the models enable it to run on chips with low memory, faster.
The bitnets marks the weight as just three values: -1, 0 and 1. By theory, which makes them much more memory- and computing-sight than most models today.
Microsoft researchers say that Bitnet B 1.58 2B4 is the first bitnet with 2 billion parameters, “parameters” are basically “weight” synonym. The equivalent of about 33 million books is trained on the data set of 4 trillion token, By an estimate – BitNet B.58 2B4 has surpassed the traditional -shaped models of similar size, researchers claim.
BitNet B 1.58 2B4 rivals cannot clean the floor with 2 billion-parameter models, but it is apparently its own. According to the researchers test, the model leaves the Mater Lama 3.2 1B, Google’s Jemma 3B, and Alibaba Queen 2.5 BSM 8K (compilation of grade-school-level mathematics) and packers (which tests the logic of physical commonsense).
Perhaps more impressively, Bitnet B 1.58 2B4’s size is faster than other models of size – in some cases when using a fraction of memory twice, twice the speed.
However there is a catch.
To achieve that performance, Microsoft’s custom framework, Bitnet.PPP need to be used, which only works with certain hardware at the moment. GPU is missing from the list of supported chips, which prioritize the AI infrastructure landscape.
This is to say that Beatnets can keep a promise, especially for resource-conferred devices. But the consistency is – and will probably be – a big sticking point.