BitNet b1.58 (1-bit LLM)
Every weight is ternary {-1, 0, 1}; paper claims parity with full-precision transformers at the same size and tokens with large efficiency gains.
Proposed training LLMs natively at ~1.58 bits per weight, replacing multiplications with additions and cutting memory, latency and energy; positioned as a new scaling recipe and a case for specialised hardware.
- Date
- Tuesday, 27 February 2024
- Lab
- Microsoft
- Kind
- paper
- Access
- research preview
The paper's results were at smaller scale; an open 2B model followed in April 2025.
Sources
This record was checked against its sources on 6 October 2026. How we check