Start an inquiry
A 32B doesn't belong on a 16 GB GPU.
Ours lives there.
Qwen3-32B, Gosset Packed: 3.6577 bits per weight, 100% resident on a single RTX 5080. No offloading, no spilling to system RAM, and 93.48% on the full GSM8K set.
It's Gosset Packed.
Gosset Pack № 1: Qwen3-32B
Gosset Packing rounds groups of eight weights to the nearest point of the E8 lattice, the densest way to pack in eight dimensions, described by Thorold Gosset in 1900. The result is a model file small enough to live entirely in a consumer GPU's memory, decoded directly in registers at interactive speed.
A pack is a compiled output. The encoder never ships.
Packs are released gated, to verified recipients. Public perplexity figures and full benchmark sheets are published with each pack.
Gosset Pack № 2: Qwen3-14B
The method proof, and the pack that gets daily-driven. Qwen3-14B in two builds, audited end to end: the release files are bit-identical to the checkpoint that produced these numbers, verified tensor by tensor.
| Build | Bits per weight | WikiText-2 perplexity vs BF16 |
|---|---|---|
| Qwen3-14B-Gosset-3.43BPW | 3.43 bpw | +1.89% |
| Qwen3-14B-Gosset-4.25BPW | 4.25 bpw | +1.58% |
Perplexity on the WikiText-2 test set, 2,048-token chunks, against the BF16 model in the same harness. Full-set task batteries headline the 32B flagship above; the 14B line carries the fidelity proof.
Measured, not promised
Every figure below was measured on the shipped v3 pack, on one RTX 5080, against the full official benchmark sets. No samples, no subsets.
| Benchmark | Result | Detail |
|---|---|---|
| GSM8K | 93.48% | 1,233 of 1,319 problems, full set |
| HumanEval | 86.59% | pass@1, 142 of 164 problems, full set |
| MBPP | 72.40% | pass@1, 362 of 500 problems, full set |
| WikiText-2 perplexity | +5.4% | Versus the BF16 model, long-form protocol |
| Decode speed | 38 to 40 tok/s | Fully resident, no offloading |
| GPU memory | 12.95 GiB | 100% resident on a 16 GB card |
The 16 GB problem
A 32B model at full precision needs about 65 GB. Standard 4-bit builds still need about 20 GB, so on a 16 GB card they spill into system RAM and slow to a crawl. Gosset Packing was built to close exactly that gap.
| Format | Size | On a 16 GB GPU |
|---|---|---|
| BF16 (full precision) | ~65 GB | Does not fit |
| Standard 4-bit | ~20 GB | Spills to system RAM |
| Gosset Packed, 3.6577 bpw | 14.62 GB | 100% resident |
Private packing
We pack private models for teams who need them running on hardware they already own: verified deployment and a validation report on your own evaluation, on your premises. Engagements are inquiry-only for now. Tell us what you need packed and what it has to run on: