Keep the family. Lose the repetition.
Less storage than five self-contained deployments in the measured model-and-adapter family.
Install once · one model library
Keep the models. Get your disk space back.
Connect your model folders once and stop sorting files by hand. Tensor Archive organizes models, quantizations, LoRAs and adapters into exact, lossless families, ready when your model apps need them.
- Automatic family organization
- 100% lossless
- Local and private by design
Powered by U.S. Patent-Pending Technology
Measured model family
Exact restore
5 self-contained deployments
907.3 MB
Each deployment carries its own complete package.
Tensor Archive · same versions
261.6 MB
No quantization. Every restored byte matches.
Large archives. Small RAM footprint.
Less peak packaging memory than ZipLLM on the same TA-Bench v1 source data.
Exact means exact.
Bytes differed after restoring all 50 files in the sequential-checkpoint validation.
Install it. Point it at your models. Forget the folders.
One compact library, ready on demand.
Tensor Archive groups related model files, stores their shared data once and prepares the exact version a connected runtime needs. After use, the temporary working copy can be reclaimed without losing the model.
-
01
Find and organize
Connect your model folders. Related models, quantizations, LoRAs, adapters and checkpoints become clear families.
-
02
Pack losslessly
Tensor Archive stores repeated model data once while preserving every file and every byte needed for exact recovery.
-
03
Prepare on demand
The exact archived version is materialized and handed to the connected runtime when you want to load it.
-
04
Reclaim after use
Unload safely, recover the temporary working space and keep the exact model packed for its next request.
BERT Tiny plus three adapters. Every version restored exactly.
Already using LM Studio, Ollama, FLUX or ComfyUI? Keep one compact library behind them all. Tensor Archive prepares the exact model each app requests, then safely reclaims its working copy.
Try it on your model libraryOne storage layer behind your model apps
Your apps ask. Tensor Archive prepares.
Install it once. Tensor Archive automatically organizes and losslessly compresses your models, LoRAs and adapters into one compact library. When an integrated app asks for one, the exact version is prepared on demand and compacted again after use.
Available now
Dynamic model loading for Bionic.
The new Bionic from LM Studio made us not think twice about fully integrating Tensor Archive to enable dynamic model loading: compact unused models, keep many versions organized, and use the best model for every task—hassle-free.
Organize · prepare · load · reclaim
Ollama
Available now
API-first model workflow.
Ollama runs models the way engineers expect: fast, scriptable, and API-first. Where others focus on the chat window, Ollama focuses on the model workflow. Tensor Archive belongs there: compress, organize, and load the best version for each job—automatically.
Fast · scriptable · automatic
Available now
FLUX models, checkpoints and LoRAs.
FLUX is where image models and LoRAs multiply fast—and disk space disappears faster. We integrated Tensor Archive so every checkpoint and adapter stays losslessly compressed, clearly versioned, and ready to load without the usual mess.
Lossless · versioned · ready to load
ComfyUI
Available now
Serious pipelines. One compact library.
ComfyUI runs serious visual pipelines—and serious model collections. Tensor Archive keeps those checkpoints and LoRAs compact, organized, and instantly usable, so you build workflows instead of managing folders.
Checkpoints · LoRAs · workflows
T2T distributed delivery
Compact model families move as verified objects.
Lite can receive and restore T2T model families for free. Your private archive remains local; network participation is explicit and private by design.
Measured TA-Bench run
3.2× faster packing
33.02 seconds for Tensor Archive versus 105.65 seconds for Gear CDC on the same source bytes, with exact restoration. One controlled run, not a universal speed guarantee.
Measured on the models people actually keep
Real GGUF families. Real LM Studio inference.
Two Qwen2.5 families show the storage reclaimed at compact and 7B scale. The product loop was then validated end to end in LM Studio.
Qwen2.5 0.5B
Three quantizations · Q4_K_M, Q5_K_M and Q6_K
60.8% less physical storageQwen2.5 7B
Four quantizations · nine complete GGUF shards
6.26 GB reclaimed · 24.9% lessReal LM Studio inference
Activate, register, load, generate, unload and safely evict. A loaded model cannot be evicted, and every protected shard restores byte for byte.
- Load + inference
- Safe eviction
- Exact restore
TA-Bench v1 · same source bytes
Exact restore
Physical storage on one model family
Separate TA-Bench model-family workload, not the Qwen GGUF results above. Xet uses the official xet-core v1.5.4 local reference CAS; ZipLLM uses the smallest exact-restorable closure independently verified.
Don’t trade disk pressure for memory pressure
Large models. Tiny memory footprint.
Large model files do not need large memory spikes. On the same TA-Bench v1 dataset Tensor Archive peaked at 52.4 MB, versus 1,523 MB for ZipLLM.
96.6%
less peak packaging RAM than ZipLLM
31.7 GBphysical model
→
<167 MBmaximum peak RAM
Pack106.7 MB
Verify163.1 MB
Restore166.0 MB
- 1.07 GB hard cap
- Zero swap
- Zero OOM
- Exact SHA-256
Two separate proofs. The 96.6% comparison uses the shared TA-Bench v1 corpus. The 31.7 GB run validates Tensor Archive alone; ZipLLM was not run on that large file.
Inspect the memory evidenceMac Studio storage economics
Buy less SSD. Keep more models.
Apple charges for factory storage once. Tensor Archive earns that space back across every related model family. If your family matches our measured 71.2% result, a smaller SSD tier can hold the same versions.
Without Tensor Archive
2 TB Apple SSD
+$800
With Tensor Archive
576 GB physical
1 TB Apple tier +$300
Without Tensor Archive
4 TB Apple SSD
+$1,800
With Tensor Archive
1.15 TB physical
2 TB Apple tier +$800
Without Tensor Archive
8 TB Apple SSD
+$3,800
With Tensor Archive
2.30 TB physical
4 TB Apple tier +$1,800
Lite stays free.
Try it free now
Lite automatically keeps downloaded models compact, expands them when needed and compacts them again after use. T2T delivery is free too. Pro adds offload, letting you move the model archive off your disk entirely.
Illustrative equivalent capacity, not general disk compression. The 71.2% result comes from five related self-contained model deployments; results vary by family. Apple pricing is the current U.S. Mac Studio M4 Max factory-storage ladder, checked July 27, 2026, before tax. Check Apple’s configurator ↗
The numbers answer different questions
One promise. Three separate proofs.
Storage, memory and integrity are measured independently. Every result below names its workload and comparison so the numbers never have to rely on fine print.
71.2% less storage
Five self-contained model-and-adapter deployments compared with the same family in Tensor Archive.
69.4% less than Gear CDC
Five BF16 model checkpoints on the same TA-Bench v1 source bytes, with exact restore.
96.6% less peak RAM
ZipLLM and Tensor Archive packaging the same TA-Bench v1 model family.
Integrity / exact restoration
50 / 50 files restored
0 bytes differed
SHA-verified
The public model suite also reclaimed 76.33–76.46% across BERT Tiny, DistilBERT Base and SmolLM2 135M. Results vary with how closely related the versions are.
The technical signal
Generic storage left space behind. Tensor Archive reclaimed more.
Five sequential BF16 training checkpoints. Same 4.06 GB source, same files, exact restoration. Tensor Archive used 964.5 MB less physical storage than normalized Gear CDC.
SMOLLM2-135M · 5 TRAINING CHECKPOINTS
Product signal
Sequential training family
vs Gear CDC−34.318%
Files restored50 / 50
Bytes differed0
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
ZipLLM default workspaceraw + compressed1.661 GB
Raw model files1.345 GB
Gear CDC1.080 GB
tar + zstd1.052 GB
ZipLLM minimum exact restoresmallest verified subset585 MB
Tensor Archive331 MB
Track B · complete versions
Training checkpoints
Five complete versions
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. ZipLLM's default workspace retains raw and compressed tensors; our primary comparison uses the smallest subset we verified still restores every file exactly, not an official ZipLLM garbage-collection mode.
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 public suite compares five self-contained deployments per family and reclaimed 76.33–76.46% across BERT Tiny, DistilBERT and SmolLM2. 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 recordFive sequential model checkpointsPhysical size · same scale
Source checkpoint familyFive BF16 model files
1.345 GB
Gear CDCNormalized byte-level control
1.080 GB
ZipLLMSmallest verified exact-restorable store
585.4 MB
Tensor ArchiveModel-aware exact archive
330.6 MB
69.4% less storage than Gear CDC.
43.5% less than ZipLLM's smallest verified exact-restorable store.
Private by design
Your models stay your models.
Tensor Archive runs locally. It does not need destructive quantization or a cloud upload to reclaim space, and T2T participation never turns your private model library into a public hub.
Local workspace
Your archive and model files remain on infrastructure you control.
Exact recovery
Verification and SHA receipts prove that restored files match their source.
Safe activation
A model in use cannot be evicted. Unload it first, then reclaim the cache.
Private T2T delivery
Receive compact, verified model families through the distributed network while your private archive stays under your control.
Also from Tensor Archive
Want 51% of your game drive back too?
Tensor Archive Games reclaimed 51.92% of the durable storage occupied by our measured 12-title PC library—while keeping every restore exact and Steam in control.
Reclaim game drive space
Game drive / measured corpus
Exact restore
Disk reclaimed−51.92%
Restore integrity100%
Working product · Available now
Keep the models. Get your disk space back.
Connect your model folders once. Tensor Archive organizes the families, shows what you can reclaim and keeps the exact version your runtime needs ready on demand.
macOS · Windows · Linux · 100% lossless · private by design