Skip to content
Torad. LAB / 001
Talk to us
Torad Labs · small models you own runs offline

Torad Labs

Small AI models you own.

We train a model on your work and hand you the weights. It runs on your machine, offline. No meter, no per-token bill, nothing reporting back.

Start a model See the products ↓

A quarter of the size · runs offline · yours to keep

Products

Two products. One outcome: a model that is yours.

Since 2026, your own computer can run real AI. The missing piece is a model that knows your work, so we build it and you keep it.

Trainingavailable now

Torad Etch

Bring your documents and your tasks. Etch trains a small specialist that knows them cold, then hands you the weights as a file you control.

Start a model →

Runtimeships with Etch

Torad Edge

Runs your model on hardware you already own, fully offline. It serves and keeps improving the model in one program, and nothing reports back.

See Edge →

Kandi built / app coming

A festival guide running an Etch-trained model entirely on the phone, offline. The pipeline, shipped end to end.

 All apps →

The proof

We made a model four times smaller. You cannot tell the difference.

We ask the original and the shrunk model the same questions and score how alike the answers are. That score is the A score from 0 to 1.00 for how alike two models' answers are. 1.00 means identical. It measures whether the answers point the same direction, not just whether the words match.: 1.00 means identical.

How alike is the shrunk model to the original?

VALIDATED
0.9998 cosine similarity .80 1.00
Round-to-nearest0.85Round every number off. The model drifts into nonsense within a sentence.
Rotation + codebook0.91The careful static methods. You would still feel the model get duller.
TQ4 · ours0.9998The same answers, at a quarter of the size. On 5 of 5 test prompts the top answer matched the original.
31%of the memory
870 MBvs 2.7 GB in bf16
238 / 267tok/s · within 10%
Measured on Qwen3-1.7B in our 4-bit Torad Quant 4-bit, our own format. It stores each number in 4 bits instead of the usual 16, so the model takes about a quarter of the space. format: the same answers, a quarter of the size, at the same speed. Holds on other models too: Google's Gemma 4 E2B scores 0.9997.

The cloud cannot do this. Our model runs on a graphics card you can buy, so cut the connection and watch.

The same model, on a 16 GB graphics card VALIDATED

Hosted cloud online Torad Edge local
8B model on a 16 GB GPUcannot fit40 tok/s
14B model on a 16 GB GPUcannot fit21 tok/s
your data leaves the machinealwaysnever

Being honest. Every number on this wall is measured, and you can re-run each one. The claims we believe but have not proven yet are named on the technology page, never here.

How it works

Train it once. Run it on hardware you own.

domain CPT SFT RL your machine yours
Your domain goes in. A 4-bit model you own comes out, running offline.
Torad Etch → Torad Edge → The technology →

Why it matters

Intelligence you do not own can be revoked.

The AI you rent is metered, logged, and can change without warning. A model you own is none of those things.

The Right to Your Own Intelligence →

Get started

Get the model that knows your work.

Be first to know

We write when Etch and Edge ship. No noise.

runs offline · Torad Labs