Downloads · 30 days
0
bong-water-water-bong/ZAYA1-8B-1BP
ZAYA1-8B-1BP is a machine learning model from bong-water-water-bong. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Zyphra/ZAYA1-8B (Apache-2.0) converted to 1BP — the single-file model format used by the 1bit.MONSTER engine.
Downloads · 30 days
0
Access
Public
Updated Sep 14, 2026
Repo size
12.9 GB
Likes
2
Public
Click a slice to open those files.
.1bp6.3 GB · 100%
From the Hugging Face model README
Zyphra/ZAYA1-8B (Apache-2.0) converted to 1BP — the single-file model format used by the 1bit.MONSTER engine.
1BP packs everything a loader needs into one memory-mappable file: a 256-byte header
with the model config, a variable-length tensor index, and Q4NX-tiled (32×256, 4-bit
groups of 32 with bf16 scale/zero-point) weight data. No external config.json.
1203 tensors from the source model, categorized by rank:
| Kind | Count | Storage |
|---|---|---|
| Norms / biases / scales (1D) | 682 | Raw float32 — small, precision-sensitive, not quantized |
| Weight matrices (2D) | 401 | Q4NX-tiled, 4-bit (81 of them F16: routers, embeddings) |
| MoE expert stacks (3D, 16 experts each) | 120 | Each expert's slice independently Q4NX-tiled |
Total 8.84B parameters, converted from the source GGUF exactly
(zaya1-8b-Q4_K_M.gguf, 5,567,131,712 bytes) — this file is 6,321,669,632 bytes
(6.0 GB), 1BP format v4, arch MOE, quant Q4NX, sha256
41f8fb517e06a6d3a528af48a5675c82496886592db31189fe3bbc2e00f56e00.
Header: H=2048 L=40 NH=8 NKV=2 HD=128 NE=16 vocab_size=262272, and
token_embd.weight is [262272, 2048] — i.e. the declared vocab and the embedding
table agree (see the history below for why that is worth stating).
ZAYA1-8B.htok — the matching BPE tokenizer written by the same conversion:
262147 tokens, 514906 merges, 10 specials. The 1BP file carries tensors only, so
the tokenizer ships beside it; the engine looks for tokenizer.htok next to the
weights (~/.local/share/1bit-monster/weights/).
2026-07-20 — first complete upload (all 1D/3D tensors captured; a previous
attempt had only ~773M of the 8.84B parameters because the converter dropped every
norm and every MoE expert tensor — see the feat/onebp-moe-experts history in the
old 1bit-systems repo).
2026-09-14 — layout fix (this revision). The 2026-07-20 file was format v1
and had been written by a converter that enumerated each attention/MoE sub-block as
its own blk.N: 80 block indices / 1283 tensors, with a header claiming
vocab_size=262272 while token_embd.weight held only 262147 rows — an
internally inconsistent file that the current engine's loader refuses (the
zaya_init_onebp: model embed size 536877056 != expected 537133056 failure), and
one whose block layout does not match the Zaya graph (the reference config says
num_hidden_layers: 40).
Re-converted from the same GGUF with the current converter
(gguf_to_onebp … --q4nx), giving format v4, 40 block indices / 1203 tensors,
and a self-consistent header (vocab_size 262272 = embedding rows 262272). Every
tensor name, shape, tiled byte size and per-tensor quant in this revision matches
the engine-validated zaya-q4k.1bp produced from the same GGUF by the same
converter (1203/1203, diff-clean), so this is the layout the engine actually serves.
Superseded revision (kept in git history): 6,577,168,970 bytes, LFS oid
5abef9b00ab4d349….
git clone https://github.com/1bit-MONSTER/1bit-MONSTER
cd 1bit-MONSTER
cmake -B build && cmake --build build --target onebin -j$(nproc)
cmake --build build --target npu_engine_universal -j$(nproc) # the NPU lane (needs XRT)
export NPU_ENGINE_BIN=$PWD/build/npu_engine_universal
./build/1bit zaya -m /path/to/ZAYA1-8B.1bp
Zaya is a hybrid SSM/MoE model, so the dense GPU lane does not apply — NPU
(npu_engine_universal, no FastFlowLM required) or the Zaya HIP path are the
supported lanes. See docs/wiki/npu-architecture.md for the lane map.
Base model: Zyphra/ZAYA1-8B, 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.