Model-native storage and versioning
Stop storing the same model five times. Use up to 71% less space.
Keep every checkpoint and adapter. Exact recovery. Private by design.
- 01 Working product
- 02 Public benchmarks
- 03 Private deployments
One number to remember
Measured
5 standalone deployments907.3 MB
Tensor Archive
261.6 MB
−71.2%
Why this exists
Every fine-tune spawns more versions. Most teams still save full copies.
That burns disk, bandwidth and time—for weights you already have. Tensor Archive stores a model family as a family, so you keep every useful version without paying for the same base again and again.
The technical signal
Byte-level deduplication stopped. Tensor relationships kept going.
Five sequential BF16 checkpoints with AdamW state. Same 4.06 GB source family, same files, exact restoration. Tensor Archive stored 964.5 MB fewer physical bytes than normalized Gear CDC.
SMOLLM2-135M · 5 CHECKPOINTS · ADAMW
Product signal
Sequential training family
vs Gear CDC−34.318%
Files restored50 / 50
Tensors verified4,080
Replay restore ratio0.95×
Restore ratio is measured against the exact fixed-chunk control; below 1× means no observed slowdown in this replay. This is one controlled workload, not a universal claim.
TA-Bench v1 · direct comparison
Same source bytes. Exact restores. No borrowed numbers.
Five sequential BF16 checkpoints from one frozen training run. Every scored system receives the same files; every bar is measured physical storage, including the metadata required to restore them.
Track A · model only
Model family
5 model files
1.345 GB source
Raw files1.345 GB
tar + zstd1.052 GB
Gear CDC1.080 GB
ZipLLM · official store1.661 GB
ZipLLM · audited closure585 MB
Tensor Archive331 MB
Track B · full state
Training checkpoints
Weights + AdamW state
4.060 GB source
Raw files4.060 GB
tar + zstd2.758 GB
Gear CDC2.810 GB
Tensor Archive1.846 GB
Strict scope
FM-Delta is N/A because its official codec does not accept this BF16 corpus. DCP belongs in a native semantic-state track. Gear CDC is a normalized byte-level control, not a claim about the hosted Xet service.
Storage run 001 · corpus manifest bfa502…3ce81 · every scored restore
byte-exact. Full Tensor Archive timing repetitions and independent reproduction remain pending.
No benchmark theatre
One number attracts attention. Every baseline stays visible.
The 71.2% result compares self-contained deployments that each carry their own base model; it is separate from TA-Bench. The direct result above compares Tensor Archive, ZipLLM, zstd and CDC on one frozen sequential model family.
Inspect the public metric recordSame base + four adaptersPhysical size
Self-contained deploymentsBase repeated five times
907.3 MB
Separate filesOne base + adapters, no archive
323.3 MB
Gear CDCNormalized 1 MiB target chunks
261.7 MB
Tensor ArchivePacked family catalog
261.6 MB
Tensor Archive saves 19.1% vs separate files.
On this adapter catalog, only 0.027% vs Gear CDC.
What you get
Built for how teams actually work.
Base model + LoRAs. Checkpoint sequences. Training snapshots with optimizer state. Tensor Archive is the data layer for that reality—not another folder of duplicated files.
01 / ingestCheckpointsmodels · adapters · optimizer state
02 / understandTensor Archivefamily graph · exact references · catalog
03 / restoreAny versionrandom access · byte-exact promotion
Less storage
Model + four adapters: 71% less space than self-contained deployments.
Harder proof
Five checkpoints + optimizer state: 34% less than a strong CDC baseline, with fast restore.
Exact, not “close enough”
Byte-for-byte recovery. SHA-verified. No destructive quantization.
Your infra, your weights
Private archive and registry. No need to hand models to a public hub.
Ship versions by name
Publish, sync and pull what changed—not the whole family every time.
Where we are
Working product. Public benchmarks. Private deployments available now.
Proven Exact round-trip, random tensor access and the first benchmark.
Available Private deployment for real checkpoint and adapter families. Commercial onboarding is open.
Next Independent reproduction, PyTorch DCP integration and broader benchmarks.
Available now
Keep every model version. Store less.
Stop duplicating the same weights. Keep every recovery point privately and exactly.
Book a deploy conversation hello@tensorarchive.ai