TL;DR
A class of highly efficient large language models where internal weights are restricted to extremely low precision, typically using binary or ternary values.
One-bit large language models replace the standard sixteen-bit floating-point weights used in conventional systems with extremely quantized values, often restricted to just negative one, zero, and positive one. This structural change replaces resource-heavy matrix multiplication with simple, energy-efficient addition operations during inference. Consequently, these models significantly decrease memory consumption and latency while demonstrating capabilities competitive with full-precision alternatives at scale.
Why this matters for your business
This approach enables high-performance AI reasoning to run directly on consumer-grade hardware and edge devices, drastically lowering operational costs.
