Microsoft researchers have created the largest-scale 1-bit AI model to date, called BitNet b1.58 2B4T. This “bitnet” is designed to run on lightweight hardware and can be used on CPUs, including Apple’s M2. By quantizing weights into just three values (-1, 0, and 1), bitnets are significantly more memory- and computing-efficient than traditional models.
The new model boasts 2 billion parameters, trained on a massive dataset of 4 trillion tokens, which is equivalent to about 33 million books. It outperforms similar-sized traditional models, surpassing Meta’s Llama 3.2 1B and Google’s Gemma 3 1B on benchmarks such as GSM8K and PIQA.
BitNet b1.58 2B4T also offers impressive speed gains, running twice as fast while using a fraction of the memory required by other models of its size. However, achieving this performance requires Microsoft’s custom framework, bitnet.cpp, which only supports certain hardware configurations, excluding GPUs from the list of supported chips.
Despite compatibility issues, researchers believe that bitnets hold promise for resource-constrained devices, where efficiency is crucial.
Source: https://techcrunch.com/2025/04/16/microsoft-researchers-say-theyve-developed-a-hyper-efficient-ai-model-that-can-run-on-cpus