Start with a real public family or your own folder.
On an empty Overview, select Build real public family. Tensor Archive downloads a fixed BERT Tiny sample and opens the measured result. The reference run represented 72,640,943 logical bytes in 17,178,128 stored bytes: 76.35% less space. Results can change if the published sample changes in a later release.
The sample downloads only declared files from public Hugging Face repositories; it never executes model code or weights. If you are offline, choose Use offline synthetic demo. The separate Google BERT Tiny preset remains a single-checkpoint compatibility check and is expected to show limited savings.
-
1
Build the public family
Choose Build real public family to open a measured sample and review its storage result.
-
2
Or select Add model family
From Overview, choose Local folder, then use Choose folder.
-
3
Analyze before writing
Select Analyze source. Review the detected type, source size and expected result.
-
4
Name and archive
Give the family a clear archive name and select Archive folder. Progress appears as a local job.
-
5
Verify, then test a restore
Open the family, run Verify, restore a version, and compare or open it before deleting any source copy.
Archiving creates a managed copy; it does not move or delete your source. Keep the original until Verify succeeds and you have tested an exact restore.
Understand what each storage number means.
Tensor Archive separates the size your versions represent from the bytes physically required to keep them.
The total size if each archived version were kept as its own self-contained copy.
The bytes currently occupied by the family inside Tensor Archive's local storage.
The difference between the full represented size and the disk space currently used.
The extra disk space used when a new version is added to the archive.
76.35% is the public BERT Tiny demo result. A separate larger reference family measured 71.2%. A same-shape public suite also measured DistilBERT at 76.46% and SmolLM2-135M at 76.33%, with exact restoration in all twelve cases. None is a promise for every folder: families that share substantial content can save a lot; unrelated or already compressed files may save little.
Keep versions from the same project together.
Clear grouping makes storage results easier to understand and each version easier to find when you need it.
Checkpoints and model directories
Choose a folder using the native system picker. Tensor Archive receives temporary access to that selection, checks it locally and rejects unsafe links or special files.
Attach a LoRA to its archived base
Add the main checkpoint and its adapters to the same family. Tensor Archive guides you when it needs confirmation before continuing.
Resolve a public repository revision
Enter owner/model and a branch, tag or commit. Analyze reads the public file inventory and may inspect a small adapter configuration; it does not download or execute model weights. Archive then resolves an immutable commit and stores the complete public snapshot locally. The Google BERT Tiny preset offers a small official first source.
Use one family for genuinely related models
Group a base, its adapters and its real revisions together. Separate unrelated architectures or projects so that each archive stays clear.
Inspect, prove and recover what you saved.
Creates a durable receipt showing that the saved version can be recovered exactly. It does not assess model quality or accuracy.
Creates a byte-for-byte output in the app's local restore area.
When the job completes, save the copy to Downloads without overwriting an existing file.
Review its files, sizes and checksums before restoring it.
Archive, verify and recovery actions run locally with progress, status and completion receipts.
Review managed storage and generated recovery outputs. Copies already saved to Downloads are outside the app and are not counted.
Removes that family's managed data while leaving existing copies in Downloads untouched.
Delete originals only after you prove recovery.
Tensor Archive deliberately does not remove source data. That final storage decision stays with you.
1Archive the folder and wait for successful completion.
2Verify the archived version and retain its receipt.
3Restore it to a separate output and confirm your workflow can use it.
4Back it up externally if loss of this Mac would matter.
5Delete the old source manually only when you are satisfied with those checks.
Local archiving saves disk space, but it does not protect against disk failure, theft or accidental removal of the application data directory. Keep another copy of irreplaceable model data.
Where Tensor Archive is most useful today.
- A base model with several LoRAs or adapters.
- Related checkpoints and training revisions.
- Comparing the disk space used by related versions.
- Freeing local SSD space after verified recovery tests.
- Auditing archived files and recovery receipts.
- Exact offline restores from a local archive.
- A training or inference accelerator.
- Quantization, pruning or a model conversion tool.
- A way to improve model quality or accuracy.
- A guaranteed 76% saving for every collection.
- A cloud backup or multi-device sync service in this release.
- A replacement for an independent backup of important data.
Tensor Archive Local supports Apple silicon Macs running macOS 13 or later, x64 PCs running Windows 10 22H2 or Windows 11 and x64 PCs running Ubuntu 24.04. Integrated workflows for LM Studio, Ollama, FLUX and ComfyUI are available. Public Hugging Face repositories and local folders are also available. Private repository authentication, S3, cloud sync and team libraries are not available yet.
Your local models stay local.
No account is required
Local folders and archive contents are not uploaded to Tensor Archive. Network access is used when you explicitly choose Hugging Face and when the app checks tensorarchive.ai for product updates. Anonymous product analytics are enabled by default and can be disabled under Settings.
Usage signals, never model data
Tensor Archive records app opens, active time, the operation used, its outcome, app version, platform and update results under a separate random analytics identifier. It never sends model weights or archive contents, model, repository or file names, local paths, hashes, email or license identity. Raw events are kept for at most 90 days. Turning analytics off requests deletion of this installation's events.
One private application directory
Archived content, receipts and generated outputs live under
~/Library/Application Support/ai.tensorarchive.local/ on
macOS, %LOCALAPPDATA%\ai.tensorarchive.local on Windows or
~/.local/share/ai.tensorarchive.local/ on Linux. Removing
the matching directory removes the local archive too.
A protected 14-day trial or Local Pro
Each installation can start one free 14-day trial without a card or account. Viewer, Verify and Restore remain available afterward; creating or extending families requires an active trial or Local Pro and an internet connection. Open License in the app to buy Local Pro for $199 USD through Stripe Checkout and activate up to three machines. Stripe processes the live payment; Tensor Archive never receives card details.
Verified update metadata
The app validates the signed update manifest and exact SHA-256 before offering a new version. If an older installation cannot update automatically, download the current installer from the official Tensor Archive page and install it once.