The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

TL;DR

This paper introduces BitNet b1.58, a 1.58-bit ternary LLM that matches full-precision models in performance while significantly reducing costs, and establishes new scaling laws and hardware opportunities for 1-bit LLMs.

Contribution

The paper presents a novel 1.58-bit ternary LLM, BitNet b1.58, demonstrating high performance and cost efficiency, and proposes new training scaling laws and hardware design directions.

Findings

BitNet b1.58 matches full-precision LLM performance.

1.58-bit models are more cost-effective in latency, memory, and energy.

Defines new scaling laws for high-performance, low-bit LLMs.

Abstract

Recent research, such as BitNet, is paving the way for a new era of 1-bit Large Language Models (LLMs). In this work, we introduce a 1-bit LLM variant, namely BitNet b1.58, in which every single parameter (or weight) of the LLM is ternary {-1, 0, 1}. It matches the full-precision (i.e., FP16 or BF16) Transformer LLM with the same model size and training tokens in terms of both perplexity and end-task performance, while being significantly more cost-effective in terms of latency, memory, throughput, and energy consumption. More profoundly, the 1.58-bit LLM defines a new scaling law and recipe for training new generations of LLMs that are both high-performance and cost-effective. Furthermore, it enables a new computation paradigm and opens the door for designing specific hardware optimized for 1-bit LLMs.