Microsoft Research presents BitNet 1.58: The 1-bit revolution

The Era of 1-bit LLMs

Almost simultaneously with Google's announcement, Microsoft Research published its work on BitNet 1.58. The concept is as simple as it is revolutionary: instead of using floating-point parameters (16-bit or 8-bit), BitNet uses only three states for its weights: -1, 0, and 1 (roughly 1.58 bits per parameter).

Why does it matter?

By replacing the majority of multiplication operations with simple additions, BitNet 1.58 drastically reduces memory and energy consumption while maintaining performances comparable to classic models of the same size.

Edge AI and Gaming

This architecture paves the way for running ultra-powerful language models directly on consumer devices (smartphones, consoles). In the context of Vibe Hosting and Novoid, the arrival of models like BitNet could allow deploying high-performance local AIs without requiring prohibitively expensive GPU clusters.