BitNet b1.58 2B4T

BitNet b1.58 2B4T is Microsoft Research's proof that LLMs don't need floating-point weights: a 2B-parameter model trained natively with ternary weights (-1, 0, 1), i.e. 1.58 bits per weight, on 4 trillion tokens. Because matrix multiplication becomes addition and subtraction, the model runs efficiently on plain CPUs via the open-source bitnet.cpp runtime — reporting 2-6x speedups and up to 82% energy reduction on x86 — while matching full-precision 2B-class models on benchmarks.

This is a research artifact, not a product: 2B parameters won't rival modern chat models on hard tasks, and the ecosystem (quant formats, serving tooling) is built around bitnet.cpp rather than the standard vLLM/TGI stack. But as the best-trained 1-bit model available (4T tokens dwarfs the 100B-token community attempts), it is the reference point for the entire extreme-quantization direction and a genuinely practical option for offline, battery-powered inference.