- Microsoft debuts BitNet b1.58 2B4T, a 2-billion-parameter AI model optimized to run efficiently on CPUs, including Apple’s M2 chip.
- Outperforms rivals like Meta’s Llama and Google’s Gemma on key benchmarks while using significantly less memory.
- Requires Microsoft’s custom framework, bitnet.cpp, which currently lacks GPU support, limiting broader compatibility.
Microsoft has introduced a new artificial intelligence model designed to deliver high performance on standard CPUs, a significant departure from the GPU-heavy infrastructure that powers most modern AI systems. The model, dubbed BitNet b1.58 2B4T, is a “bitnet” — a type of AI model that uses extreme quantization to reduce memory and computing requirements. Unlike typical models that rely on GPUs, BitNet b1.58 2B4T is built to run efficiently on conventional processors, including Apple’s M2 chip.
Bitnets work by compressing model weights into just three values: -1, 0, and 1. This radical simplification allows the model to function on lightweight hardware with significantly less power and memory. Microsoft’s new bitnet sets a record as the first of its kind with two billion parameters, all trained on a massive dataset of four trillion tokens — roughly equivalent to the content of 33 million books.
According to Microsoft’s internal benchmarks, BitNet b1.58 2B4T holds its ground against established models of similar size. It reportedly outperforms Meta’s Llama 3.2 1B, Google’s Gemma 3 1B, and Alibaba’s Qwen 2.5 1.5B on tests such as GSM8K, which covers math problems at a grade-school level, and PIQA, which evaluates physical commonsense reasoning.
Beyond performance, the standout feature of BitNet b1.58 2B4T is its efficiency. The model is said to run at nearly twice the speed of comparable models while using a fraction of the memory. This could open the door for AI capabilities on devices previously considered underpowered for such tasks, including laptops, tablets, and edge computing systems.
However, there are limitations. The model’s performance relies on Microsoft’s own software framework, bitnet.cpp, which currently supports only specific hardware types. Crucially, GPUs — the backbone of most AI development — aren’t yet supported. While bitnets offer a glimpse into a more accessible and efficient AI future, widespread adoption will depend on expanding compatibility across diverse hardware platforms.
