BitNet b1.58 Brings 100B-Parameter AI Inference to Ordinary CPUs

A ternary neural network architecture called BitNet b1.58 reached broad scientific adoption in late September 2026, marking a significant shift in how large AI models are run. The approach encodes neural network weights using only three discrete values — −1, 0, and +1 — requiring roughly 1.58 bits per parameter, which eliminates the need for floating-point matrix multiplication during inference. This means computations reduce to simple integer additions and subtractions, drastically cutting hardware demands. The open-source framework bitnet.cpp has since been stabilized to support mainstream CPUs including Intel Xeon, AMD EPYC, Apple Silicon, AWS Graviton, and edge chips like Snapdragon and RISC-V. Models with up to 100 billion parameters can now run on commercial servers without expensive GPU clusters, potentially democratizing access to large-scale AI inference.