Anyone can shrink a model. We prove it still behaves.
Smaller, faster inference is table stakes — free tools already get ~30% losslessly.
The hard part, the part your risk team can't sign off on, is proving the smaller model
still behaves. We hand you that proof: a receipt you can check — bit-exact on certified builds.
Compressed — fewer parameters, less hardware, lower bill.
Understandable
You learn the safe compression depth before you commit.
Assured
The certificate: preserved behavior, not “close enough.”
Verified
Every number tied to a logged, reproducible experiment.
Measured, on real hardware · 2026
1.37×
faster with zero output change — certified slice of a 1.3B model, A100 GPUs
1.11×
instant on an open 7B model — no retraining, perplexity within 0.3% of dense
−1 GPU
the fit dividend, computed from measured ratios: a 70B-class model from 3 GPUs to 2; 34B-class onto one (projected — deployment at this scale is roadmap)
Every figure ties to a logged, reproducible experiment. The full result set lives in the vault.
The certificate
Every other tool ships a promise — “almost as good, trust us.”
We ship a receipt: what you deploy reproduces the reference we certified, proven by
direct test on real hardware — bit-for-bit on certified builds, within a measured
tolerance across other kernels. It is the difference between “we pruned it and it seems fine”
and “here is the identity proof” — for your risk team, your auditor, your customers.
(Not to be confused with commodity lossless repacking — same model, ~30%
fewer bytes, no faster; useful, and increasingly free.)
Why you can trust the numbers
Every claim on this site was scored against a
prediction we froze before we ran the experiment — and we publish the misses next to the
hits. That is the whole difference between a benchmark and a receipt.
PRE-REGISTERED
Frozen, then run
Predictions and their falsifiers are committed — with a hash — before the experiment. No moving the goalposts after the data lands.
NULLS PUBLISHED
Misses on the record
Failed experiments sit next to the ones that worked. This month alone we retired, in the open: from-scratch training efficiency, an ordering “moat,” and a hoped-for recovery cliff.
TRACEABLE
Dataset + commit
Every number ties to a dataset and a commit hash — reproducible on request, not asserted on a slide.
That is not marketing discipline — it is the product. A certificate is only worth as much as the lab behind it.
The method — five moves
01
MEASURE
Profile the model on a representative workload.
02
ORDER
Rank its internal capacity so the least-used parts come off first.
03
NARROW
Remove the least-used capacity to the measured safe depth — before the quality cliff.
04
RECOVER
A short recovery pass restores the compressed model's behavior toward the target quality.
05
CERTIFY
Verify the compressed model against the reference and put the numbers on the record.
Audits survey the whole method space — our ordering, the standard
pruning baselines, and quantization stacking — at matched budgets. You see every option priced, not one vendor's favorite.
Who it's for
Teams serving open models (Llama, Mistral, Qwen) at volume — every one overpays for inference today.
Real-time & edge vision — live video is compute-bound; smaller models mean more streams per device, lower latency, and a smaller box per site.
Regulated & safety-critical buyers — finance, health, industrial — where a compressed model needs a proof, not a promise.
Why SUAV.AI
Built and run by one founder on a bootstrap budget — billion-parameter
experiments on rented A100s, local models on Apple silicon, a filed patent (2026), and every result
on the record. So when we say 1.37× and bit-exact, those words mean exactly what they say.