Microsoft's new BitNet b1.58 2B4T model with 2 billion parameters performs significantly better than competitors of the same size, while using considerably less memory.
Researchers from Microsoft have just announced the successful development of the largest 1-bit AI model (also known as "bitnet") to date.
The model named BitNet b1.58 2B4T was released under the MIT license and can run on standard CPUs, including Apple M2 chips.
An Efficient Revolution for Lightweight AI
Bitnets are compressed AI models that can run on lightweight hardware. While current standard models typically require large bit quantities to represent weights, bitnets quantize weights down to just three values: -1, 0, and 1. This helps bitnets save significantly more memory and computational resources compared to most current models.
BitNet b1.58 2B4T is Microsoft's first bitnet model with 2 billion parameters. The model was trained on a massive dataset of 4 trillion tokens – equivalent to approximately 33 million books. According to Microsoft's announcement, this model outperforms traditional models of similar size.
In performance tests, BitNet b1.58 2B4T surpassed Meta's Llama 3.2 1B, Google's Gemma 3 1B, and Alibaba's Qwen 2.5 1.5B on several important benchmarks like GSM8K (elementary school problem set) and PIQA (testing physical world reasoning capabilities).
Particularly impressive is BitNet b1.58 2B4T's speed, which in some cases is twice as fast as models of the same size, while using only a small fraction of the memory compared to competitors.
However, to achieve optimal performance, BitNet b1.58 2B4T requires using Microsoft's bitnet.cpp framework, which is currently only compatible with certain hardware. Notably, this list does not include GPUs – the chip type currently dominating AI infrastructure.
Bitnets seem very promising for devices with limited resources, but compatibility issues remain a significant barrier and may continue to exist in the near future.
