Lossless model weights, about 1.5× smaller, unpacked on the graphics card.

`.shn5` is a file format for language-model weights. It is **lossless**: every byte that comes out is the byte thatwent in, so a model run from a `.shn5` gives the same numbers, bit for bit, as the same model uncompressed. Modelsstored at full precision (BF16 or F16) come out about **1.5× smaller** (measured 1.51× on BF16, 1.48× on F16), ondisk, in memory and across the PCIe bus. Already-quantised models are stored as they come. The compressed weights are unpacked by the graphics card itself, straight into the buffers the model runs from, farfaster than the PCIe bus can deliver them. So a model loads from a `.shn5` as fast as from its uncompressed twin, and amixture-of-experts model that is bigger than the graphics card can keep its experts compressed in system memory andbring only the ones each word needs to the card. On one 16 GB RTX 4080, Qwen3-30B-A3B in full BF16 precision (61 GB,40.5 GB as a `.shn5`) writes at 16.2 tokens a second, against 9.5–10.4 for llama.cpp's expert-offload mode on the samemachine, with answers identical to running the model whole. The name comes from Claude Shannon: *shn* for the limit the format works against.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24235770
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Lossless model weights, about 1.5× smaller, unpacked on the graphics card.

Michael Ricky Neal
Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
article

Lossless model weights, about 1.5× smaller, unpacked on the graphics card.

Michael Ricky Neal
article en

Abstract

`.shn5` is a file format for language-model weights. It is **lossless**: every byte that comes out is the byte thatwent in, so a model run from a `.shn5` gives the same numbers, bit for bit, as the same model uncompressed. Modelsstored at full precision (BF16 or F16) come out about **1.5× smaller** (measured 1.51× on BF16, 1.48× on F16), ondisk, in memory and across the PCIe bus. Already-quantised models are stored as they come. The compressed weights are unpacked by the graphics card itself, straight into the buffers the model runs from, farfaster than the PCIe bus can deliver them. So a model loads from a `.shn5` as fast as from its uncompressed twin, and amixture-of-experts model that is bigger than the graphics card can keep its experts compressed in system memory andbring only the ones each word needs to the card. On one 16 GB RTX 4080, Qwen3-30B-A3B in full BF16 precision (61 GB,40.5 GB as a `.shn5`) writes at 16.2 tokens a second, against 9.5–10.4 for llama.cpp's expert-offload mode on the samemachine, with answers identical to running the model whole. The name comes from Claude Shannon: *shn* for the limit the format works against.

Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon