Skip to main content
Dits LogoDits for AI
Storage & sync

What dedup actually saves

Content-addressed dedup pays off hugely on some AI workloads and barely on others. We show both — including the case today's engine handles poorly.

Typical savings
56–76%
strong-fit workloads
Sync transfer
Delta only
changed chunks, resumable
Integrity
BLAKE3
verified on every read

By workload

Logical size is what you think you have; stored is unique bytes after dedup.

WorkloadLogicalStoredSaved
Fine-tune over a base model
7B base + 6 fine-tune checkpoints
Strong fit		182 GB44 GB76%
Quantized variant set
fp16 + int8 + int4 of one model
Strong fit		39 GB17 GB56%
Dataset snapshots
5 versions, additive + light edits
Strong fit		1.2 TB310 GB74%
Consecutive training checkpoints
10 steps, diffuse weight updates
Hard case (needs L2)		260 GB232 GB11%
Modeled, not measured. The figures above are projections from the dedup behavior of the content-addressed engine applied to representative artifact shapes — not published benchmark runs. The consecutive-checkpoint row is deliberately included to show the honest weak spot that tensor-aware chunking (L2) is designed to close. We will replace these with measured numbers as the AI engine profile lands.

Want this on your workloads?

Tell us your checkpoint cadence and artifact sizes — we'll model the savings with you.

Request early access