Native binary wire format for ternary weights — NativeTernary

[sm45118]

Hi BitNet team,
I am an independent researcher and have published a paper proposing a native binary wire format specifically designed for ternary weights {-1, 0, +1}.
The problem: BitNet b1.58 operates natively in ternary but stores weights in GGUF — a container format designed for floating-point data. There is no native wire format purpose-built for ternary weights.
NativeTernary solves this:

Each weight maps to exactly one 2-bit pair: {00}=-1, {01}=0, {10}=+1
Layer boundaries encoded as {11} (2 bits), tensor boundaries as {1111} (4 bits), model section boundaries as {111111} (6 bits) — built directly into the stream
Decoder: 2-bit register + level counter. 10 lines of code. Zero lookup tables
Stateless and resilient to bitstream corruption

Links:

arXiv: https://arxiv.org/abs/2604.03336
GitHub: https://github.com/sm45118/nativeternary

C reference implementation is in progress. Would genuinely value technical feedback from the team, and happy to discuss integration.
Maharshi Savdhariya
IIT Bombay | IIM Udaipur

[sm45118]

Update: C reference implementation is now complete with benchmark results.

NativeTernary vs GGUF for ternary weight storage at 1B parameter scale:

Format	Bits/weight	Size (1B weights)	vs NativeTernary
NativeTernary	2.000	250 MB	—
GGUF Q2_K	2.625	329 MB	1.31x larger
GGUF Q4_0	4.500	563 MB	2.25x larger
GGUF int8	8.000	1,001 MB	4.0x larger

Encode: 47–69 MB/s. Decode: 35–45 MB/s. Zero malloc, C99, ARM Cortex-M0 compatible.

GitHub: https://github.com/sm45118/nativeternary
arXiv: https://arxiv.org/abs/2604.03336

Note: GGUF sizes are estimated based on 2.625 bpw for Q2_K and standard
tensor header overhead. Synthetic uniform {-1,0,+1} weight arrays used —
real BitNet model weights may vary slightly.

Happy to discuss integration with the BitNet storage format.

[sm45118]

Updated: arXiv v2 now live with real BitNet b1.58 2B4T architecture analysis. NativeTernary framing overhead is 460x smaller than GGUF tensor headers (91 bytes vs ~42KB across 170 tensors). https://arxiv.org/abs/2604.03336