NativeTernary: A Self-Delimiting Binary Encoding with Unary Run-Length Hierarchy Markers for Ternary Neural Network Weights, Structured Data, and General Computing Infrastructure

TL;DR

NativeTernary introduces a compact, efficient binary encoding for ternary neural network weights, significantly reducing storage size and overhead compared to existing formats, with fast encoding and decoding on standard hardware.

Contribution

It presents a novel self-delimiting binary encoding with run-length markers specifically designed for ternary weights, enabling smaller and more resilient model storage.

Findings

Encodes ternary weights at exactly 2 bits per weight, 1.31x smaller than GGUF Q2_K.

Reduces tensor header overhead by 460x compared to GGUF.

Achieves encoding throughput of 47-69 MB/s and decoding of 35-45 MB/s.

Abstract

BitNet b1.58 (Ma et al., 2024) demonstrates that large language models can operate entirely on ternary weights {-1, 0, +1}, yet no native binary wire format exists for such models. NativeTernary closes this gap. Benchmarked against GGUF on the real BitNet b1.58 2B4T architecture (24 layers, ~170 tensors, 2B parameters): NativeTernary encodes ternary weights at exactly 2.000 bits per weight -- 1.31x smaller than GGUF Q2_K and 4.0x smaller than GGUF int8 -- while reducing boundary and framing overhead by 460x (91 bytes vs ~42KB of GGUF tensor headers). Encode throughput: 47--69 MB/s. Decode throughput: 35--45 MB/s on commodity hardware. The decoder is a 10-line stateless state machine resilient to bitstream corruption.