One private library for every local AI model

All your models.
Organized. Ready. Far less disk.

Tensor Archive finds the models scattered across LM Studio, Ollama, ComfyUI and your folders, then groups related versions into one private library. See what fits, compare models on your own hardware and keep every checkpoint, quantization, LoRA and adapter losslessly compact—ready when an app asks.

Free Download See it work

The first scan is read-only. Nothing is moved, uploaded or deleted.

Automatic local discovery

Scan complete

FOUND ON THIS COMPUTER

Your models, already understood.

3 connected apps

LM Studio connected

Ollama connected

ComfyUI connected

Qwen 2.5 7BQ4 · Q5 · Q6 quantizations

LM Studio 4 versions grouped

FLUX.1 familyBase model · checkpoints · LoRAs

ComfyUI 8 items grouped

Llama 3 familyRelated local variants

Ollama 3 versions grouped

01See the whole libraryAcross apps, folders and attached drives.

02Choose what fitsBefore downloading weights you cannot run.

03Use models normallyYour existing apps receive normal working files.

04Reclaim the copiesKeep exact versions without keeping every repeat.

What happens after you install it

Install it once. Tensor Archive takes it from there.

Discovery is read-only. Reclaiming space happens only after exact recovery has been proved and you choose to proceed.

It maps the library you already own.

Known LM Studio, Ollama and ComfyUI locations appear automatically. Full Scan can include other folders and attached drives.

Discovery reads model format headers and changes nothing.

It understands what belongs together.

Related quantizations, checkpoints, LoRAs and adapters become one understandable model family instead of unrelated files.

You see every usable version—and where the repeated storage is.

It proves recovery before reclaiming space.

Tensor Archive stores shared data once, rebuilds a test copy and verifies it byte for byte before the original working copy can be retired.

You decide when to proceed. No destructive quantization and no “close enough.”

The model appears when an app needs it.

Tensor Archive prepares the exact working files for the requesting app. After unload, the temporary copy can be reclaimed again.

The compact archive stays ready for the next request.

Model Discovery

Know what fits before a 30 GB download.

Model Discovery checks your operating system, memory and available acceleration locally, then shows which model sizes make sense for text, image, video, voice, music and audio—before you download any weights.

1Checks your machine locally 2Reads current model metadata 3Downloads no weights on this screen

THIS MACHINEApple silicon · 32 GB memory Profile ready

RECOMMENDED MODEL PROFILES

Choose what you want to make.

Local analysis only

TEXTLanguage + multimodalComfortable fit

IMAGEGeneration + editingComfortable fit

VIDEOMotion + cinematicsHeavy run

VOICETTS + transcriptionComfortable fit

MUSICComposition + coversExtra runtime

AUDIOSound + restorationComfortable fit

Model Lab

Benchmarks narrow the field. Your machine makes the final call.

Model Lab is a local comparison workspace. Ask a normal question, run a controlled prompt or compare two answers side by side—with the exact model identity, runtime, prompt and measured timing kept visible.

LOCAL Nothing in the prompt or result needs to leave your computer.
REAL Runs through Tensor Runtime or a connected local runtime such as Ollama.
REPEATABLE Save the prompt, environment and result as a local receipt.

MODEL LAB / SIDE-BY-SIDE EVIDENCE Local only

CONTROLLED PROMPTExplain how the storage layer preserves exact model versions.

RUN ATensor Runtime

The archive identifies repeated tensor relationships while preserving a byte-exact reconstruction path…

Ready: 1.8 s
Decode: 42.1 tok/s
Memory: 8.4 GB

RUN BOllama

The system stores shared model data once and retains the metadata needed to rebuild each original version…

Ready: 2.4 s
Decode: 37.8 tok/s
Memory: 8.7 GB

Receipt saved on this machineprompt · models · runtimes · timings · outputs

Illustrative interface. Actual results depend on the selected model, runtime, hardware and prompt.

Available now · one library behind your model apps

One library. Every model app gets the exact version it needs.

Your runtime or creative app still performs the inference and generation. Tensor Archive handles discovery, exact files and storage behind it.

LM StudioDynamic model loading

OllamaAPI-first model lifecycle

OpenCodeLocal agent workflows

ComfyUICheckpoint + LoRA workflows

FLUXVisual model families

1Select it as usualChoose the model inside the app or workflow you already use.

2The exact version appearsTensor Archive restores, verifies and exposes the required working files.

3The app runs normallyInference, image generation or automation stays with the connected app.

4Unload and reclaimThe temporary copy can leave while the compact exact archive remains.

The storage engine underneath everything

Keep every version. Reclaim the repeated space.

In a measured base-plus-adapter family, five standalone deployments occupied 907.3 MB. Tensor Archive kept the same five deployable versions in 261.6 MB. Every restored byte matched.

Open compression benchmarks

This is the proof behind “far less disk.” The full benchmark suite includes storage comparisons, RAM usage, packing speed, baselines, methodology and exact-restore evidence.

MEASURED MODEL + ADAPTER FAMILY Exact restore

Five standalone deployments907.3 MB

Each package carries its own complete base.

Store the relationships, not five copies

Tensor Archive · same five versions261.6 MB

−71.2% No quantization. Every restored byte matches.

DISK SPACE−71.2%

Measured model-and-adapter family versus five standalone deployments.

PACKAGING RAM−96.6%

Peak packaging memory versus ZipLLM on the same TA-Bench v1 source data.

EXACT RESTORE0 bytes

Differed after restoring all 50 files in the sequential-checkpoint test.

FULL EVIDENCESee every baselineOpen benchmarks →

T2T Network · delivery · elastic storage · optional compute

Your AI library can live beyond this machine. So can the work.

T2T turns a local-first library into an elastic one. Receive verified public models, move complete protected model families off this disk to eligible T2T capacity or Tensor Archive Cloud, and recover the exact bytes automatically when an app asks for them.

Optional Mutual Compute extends the same verified path to approved model preparation and reconstruction work. The result comes back verified; your local catalog remains the authority.

Your machine stays in controlStorage and compute participation are explicit.

Your private context stays privateLocal paths, private model names and prompts are not published to the common network.

T2TLOCAL-FIRSTverified exchange

01Your libraryPrivate catalog + exact model families

02Elastic storageT2T capacity or managed Cloud

03Optional computeApproved preparation + exact reconstructionCONTROLLED ROLLOUT

04Exact resultReturned, verified and ready locally

MODEL DELIVERYReceive compact public families.

Verified objects arrive through T2T and can remain compact until you need a working version.

FULL OFFLOADTake an entire family off this disk.

Protected network or Cloud capacity keeps it recoverable; Tensor Archive restores what is missing on demand.

MUTUAL COMPUTELet eligible machines prepare what your app needs.

Approved model preparation and exact reconstruction can run on selected capacity; a verified result returns ready to use.

One install · one local model system

Keep the models. Get your disk space back.

Discover, compare, compact and use every local model from one private library. Free download for macOS, Windows and Linux.

Free Download See benchmark evidence