Your machine runs the model. The weights are yours.
Edge runs the model you trained with Etch on hardware you already own, fully offline. No per-token bill, no account, no round trip.
At a glance
A third of the memory. Parity speed. Zero egress.
Qwen3-1.7B weights: 870 MB in Torad Quant 4-bit, our own format: each number stored in 4 bits instead of 16. vs 2.7 GB in bf16 · VALIDATED · Methods
Benchmarks
It runs models the cloud software cannot fit.
An 8B model does not fit a 16 GB card on vLLM. On Edge it runs at 40 tok/s. Cut the connection and compare.
| Hosted cloud online | Torad Edge local | |
|---|---|---|
| 8B · 16 GB | cannot fit | 40 tok/s |
| 14B · 16 GB | cannot fit | 21 tok/s |
| data leaves | always | never |
Same model, same card · vLLM cannot load the 8B or 14B at 16 GB · All benchmarks
How it works
One small program runs the whole thing.
One program serves the model and applies new weights in place. About 2,500 lines of Rust plus 1,300 of CUDA.
A · prompt in, answer outB · weights update in placeC · one auditable boundary
2,500 lines of Rust + 1,300 of CUDA; vLLM is ~500,000 · VALIDATED · The megakernel →
Why local
Private, unmetered, unthrottled.
Private by construction
Prompts and answers never leave the machine. There is no server to log them.
No per-token bill
The cost is hardware you already own. There is no usage charge.
No rate limits
It is your GPU. Throughput is whatever the card gives you.
Updates drop in
Retrain with Etch and the new weights replace the old file. No redeploy.
Quality
Shrinking a model usually makes it worse. Ours gives the same answers.
We ask the full-size model and the shrunk model the same questions, then 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.: 1.00 is identical.
Similarity to the full-size original
VALIDATEDdial · .80 to 1.00needle · 0.9998, oursticks · the static methods
FAQ
Questions, answered.
What hardware does it need?
A 16 GB graphics card covers the 8B and 14B models on this page. An RTX Spark or any comparable local GPU works.
Does it ever need the internet?
No. After install, serving is fully offline. Cut the connection and it keeps answering.
Which models does it run?
The models Etch produces: small specialists in our 4-bit format, trained on your data.
How do updates work?
Retrain with Etch, drop the new weights in place, and the same program serves them. No redeploy.
Why not just use the cloud?
The cloud meters you, logs you, and changes under you. Edge is a file on your disk: same answers, none of the above.
Start
The weights stay yours. Pricing coming soon.
Tell us the model and what it needs to know. No account, no card, no obligation.