Model-zoo storage

Store the base once, every checkpoint as a verified delta.

Fine-tunes and checkpoints share most of their weights with the base model — so storing each one in full is mostly redundant. AT-1 stores the base once and every variant as a verified, addressable delta. Reconstruct any checkpoint byte-exact, fetch a single tensor without unpacking the whole model, and — on the aggressive tier — ship only what a task-accuracy certificate says still holds.

Three tiers, each with a guarantee

1.4×
Lossless
byte-exact

Every weight reconstructed bit-for-bit. The reference-delta packs each checkpoint against the base so shared weights aren't stored twice.

5.5×
Zero-loss int8
no accuracy loss

Int8 quantization of the delta with zero measured accuracy loss — same predictions, far smaller footprint per checkpoint.

19.4×
Certified
within 0.12 pts

The aggressive tier, gated by a task-accuracy certificate: SST-2 accuracy moved 91.06% → 90.94%, inside 0.12 points. You ship it only because the certificate proves the task still holds.

The task-accuracy certificate is the point

Anyone can quantize a model smaller — the risk is silently breaking it. The aggressive 19.4× tier is only claimed because we measure the task and certify it: on SST-2 the model went from 91.06% to 90.94% accuracy — a 0.12-point move. The certificate travels with the stored checkpoint, so you know the compression preserved the behaviour, not just the file size.

One command surface

Initialize a zoo on a base model, add each fine-tune as a delta, then reconstruct, verify and certify any of them on demand.

# store the base once, then each fine-tune / checkpoint as a verified delta
at1 model-zoo init base.safetensors zoo/                 # the shared reference
at1 model-zoo add  zoo/ ft-sentiment.safetensors          # stored as a delta vs base
at1 model-zoo add  zoo/ ckpt-epoch3.safetensors           # another addressable delta

at1 model-zoo get    zoo/ ft-sentiment --out out.safetensors   # reconstruct any checkpoint
at1 model-zoo verify zoo/ ft-sentiment                          # -> integrity: PASS
at1 model-zoo certify zoo/ ft-sentiment --task sst2             # -> accuracy 90.94% (Δ 0.12 pt)

Reference-delta

Shared weights are stored once; each checkpoint keeps only its difference from the base.

Verified & addressable

Every checkpoint verifies byte-exact, and a single tensor can be fetched without unpacking the whole model.

Honest tiers

XOR-delta and quantization are known techniques — the value is the verification and the task certificate, and we only claim certified tiers.

Built for

ML platform teams keeping many fine-tunes of one base · checkpoint archives across training runs · model registries that must reconstruct any version byte-exact · anyone shipping a drop-in .safetensors replacement that stays verifiable.