bong-water-water-bong/ZAYA1-74B-preview-1BP

ZAYA1-74B-preview — 1BP format

Zyphra/ZAYA1-74B-preview (Apache-2.0) converted to 1BP — the single-file model format used by 1bit.systems's inference engine.

1BP packs everything a loader needs into one memory-mappable file: a 256-byte header with model config, a variable-length tensor index, and Q4NX-tiled (32x256, 4-bit groups of 32 with bf16 scale/zero-point) weight data. MoE expert stacks are stored as num_experts independently-tiled slices concatenated back-to-back.

Architecture: Mixture-of-Experts, 24 experts per layer, 120 layers, 74.79B total parameters (~4B active per token, hence "A4B").

What's in this file

All 1923 tensors from the source model — norms, dense weights, and all 24-expert MoE stacks — losslessly categorized by rank and structurally verified against the source GGUF: 74,793,442,644 parameters, exact match.

Converted from ZAYA1PREVIEW-74B-A4B-Q4_K_M.gguf (the same source this project's own NPU benchmarks used — 17.9 tok/s on Radeon 8060S iGPU, 2026-07-03; that run is now historical/archived and no longer executes on current hardware, unrelated to this file's validity as a weight artifact).

Usage

git clone https://github.com/bong-water-water-bong/1bit-systems
cd 1bit-systems
cmake -B build/engine/npu -S engine/npu -DCMAKE_BUILD_TYPE=Release   # ONEBP_SUPPORT is on by default
cmake --build build/engine/npu --target npu_engine_universal -j8
./build/engine/npu/npu_engine_universal /path/to/ZAYA1-74B-preview.1bp --model-tag zaya1_74b_preview

Attribution

Base model: Zyphra/ZAYA1-74B-preview, Apache-2.0. This repo contains a re-quantized (Q4NX, 4-bit) derivative in a different on-disk format; the original license terms apply to the underlying weights.