NeuroSquad

NeuroSquad makes open tools that keep large language models small to store and quick to ship. The first one is SeedPack.

SeedPack

A GGUF repository ships the same weights 20–30 times over: the BF16 model plus every quantization of it. SeedPack stores the BF16 model once and grows the quants back from it on the GPU, with a bit-exact CUDA port of llama.cpp's quantizers and the repository's importance matrix. Only what differs from the published files is kept, coded with the quantizer's own probabilities. A whole repository becomes one archive smaller than its BF16 model, and every file comes back byte for byte.

Qwen3.8-27B: 472.07 GB → 45.41 GB

All of unsloth/Qwen3.8-27B-GGUF — the BF16 model in 2 shards, 24 quants from UD-IQ1_S to UD-Q8_K_XL, the multi-token-prediction layer, two vision projectors and the imatrix — is one archive of 45.41 GB: 10.4× smaller than the 33 files, and smaller than the BF16 model alone (54.66 GB).

• Bigger than the machine. The 54.66 GB BF16 model never has to fit anywhere: one RTX 5070 with 12 GB and a PC with 32 GB of RAM pack the whole repository in 23 minutes, window by window, with the GPU's memory at most 8.8 GB.
• Unsloth Dynamic quants. The 21 UD quants come from Unsloth's own pipeline, whose codes drift from plain rounding along each row, so they cannot be re-derived exactly. SeedPack still stores all of them (351.53 GB) in 8.91 GB: its coder knows where in the row every weight is, and the tensors they share are stored once.
• A missing importance matrix, recovered. Q4_0 and Q4_1 were made with an imatrix that is not in the repository. SeedPack fits it back from the quants themselves, by least squares on their block scales: 97% of the blocks then come out exactly instead of 37%, and the two files cost 98.1 MB instead of about 1 GB.

Qwen3.8-Flash-Next GSQ-RCO: 294.82 GB → 112.75 GB

ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF, the GGUFs most Qwen3.8-Flash-Next users run (Strata's default), has no BF16 model to grow its sizes from. SeedPack 0.4.0 codes its 4 sizes against each other instead: identical files and tensors once, and every other tensor as a chain of versions, each coded given the precision-weighted mean of the earlier ones. All four sizes fit in one archive of 112.75 GB, 2.6× smaller than the repository. The archive is for keeping all four sizes: one size alone is still smaller.

Releases

Each archive holds a whole repository: every file of it, restored byte for byte.

Try it

pip install seedpack
seedpack list hf://neurosquad/Qwen3.8-27B-SEEDPACK                                   # what is inside
seedpack x hf://neurosquad/Qwen3.8-27B-SEEDPACK Qwen3.8-27B-UD-Q4_K_XL.gguf -o models   # one quant

SeedPack runs on an NVIDIA GPU (Turing or newer) with a CUDA 13 driver; without one, --cpu (0.5.0, pip install "seedpack[cpu]") makes the same archives byte for byte, 10–20 times slower. Python 3.11+; the model does not have to fit in GPU memory. Source code, documentation and the roadmap: GitHub · PyPI.