Microsoft’s General Artificial Intelligence group has introduced a groundbreaking large language model (LLM) that drastically reduces computational complexity by using only three weight values: -1, 0, and 1. In simpler terms, it means that this AI model can easily run on home computers and is also open-source.
Unlike conventional models that rely on 16 or 32-bit floating-point numbers, this new ternary system slashes memory and processing demands, making it possible to run high-performance AI on basic desktop CPUs.
BitNet b1.58b
Dubbed BitNet b1.58b, the model uses just 1.58 bits per weight, representing the average number of bits needed for three values. It is trained from scratch (“natively”) rather than being compressed post-training, avoiding the performance degradation seen in many quantized models. With 2 billion parameters and a 4 trillion-token training dataset, BitNet is reportedly the first open-source model trained at this bit level on such a scale.
Tiny Memory Footprint, Huge Energy Savings
BitNet b1.58b requires only 0.4GB of memory, compared to 2–5GB for models of similar size. Its ternary architecture reduces dependence on complex multiplications, relying instead on simple addition, enabling it to consume 85–96% less energy than full-precision models. It remarkably runs at human reading speed (5–7 tokens per second) using a single CPU.
Despite the drastic compression, Microsoft claims that BitNet achieves nearly the same performance on reasoning, math, and knowledge benchmarks as full-precision models in the same size class. While independent validation is still pending, Microsoft’s tests suggest impressive efficiency with minimal trade-offs.
The underlying reasons why 1-bit training works so well at scale are still unclear. Researchers admit they don’t fully understand the mechanics behind its success, calling it an “open area” for further study.
Thanks to deepseek