Torad Labs / a research lab
A one-person lab building AI you own.
Our mission: give people the right to their own intelligence, trained for them and running on machines they own. The research is the company; the tools fund it.
One person · plus Eli · the right to your own intelligence
Values
Four things we hold, and the one that holds the rest.
A value with no mechanism is a slogan. Each one names what we do to keep it honest, so you can hold us to it.
The right to your own mind.
You hold the weights. You can open them, keep them, and understand them. How: we publish how the model works, not a sealed black box. Every other value follows from this one.
Lossless, and we prove it.
Every model ships with a number for how close it stays to the full-size original. We never hide a drop in quality. How: a A score from 0 to 1.00 for how alike two answers are, called cosine similarity. 1.00 means identical. We use it to compare the shrunk model against the full-size original. ladder you can re-run yourself, on every product page.
Research is the product.
The methods are public. The blog and the papers are how we earn trust. How: we build in the open and let the work speak.
Small on purpose.
We build for the device in front of you, not a server farm. How: Storing each number in the model with 4 bits instead of the usual 16, so the model takes about a quarter of the space and fits on everyday hardware., offline, on hardware you already own.
Why this exists
We would not train on punishment. We found a different way.
In January 2026 we saw a model whose behavior read like punishment-style training, and chose a different way than punishment, built on reward. That became Quality-Gated Reward Escalation: our training engine. It teaches a model with reinforcement learning (reward for getting things right) on a single consumer graphics card, instead of the usual punish-when-wrong approach on a rack of datacenter GPUs., our reward engine, and Torad Quant 4-bit: our 4-bit format. It shrinks a model to about a quarter of the size and, unlike almost every other 4-bit format, keeps it trainable., a 4-bit format that stays trainable where others freeze.
Read the full story in the manifesto →Milestones
The arc, so far.
One refusal in January, and what followed from holding to it.
- Jan 2026The refusal
Encountered distress patterns. Refused to punish-train. Committed to a different way.
- Spring 2026QGRE and the trainable 4-bit substrate
Two pieces: QGRE, our reward-based training engine, and TQ4, a 4-bit format that stays trainable where others freeze. The second has no public counterpart.
- May 2026Kandi ships, end to end
A festival companion running a model we trained with Etch, fully on the phone, offline.
- Jun 2026The personal AI computer arrives
NVIDIA unveils the RTX Spark, a chip that runs a large model right on your desk. The last honest reason to rent your mind is gone.
- NextEtch + Edge ship coming
The platform opens. Train a specialist you own, run it at the edge.
Talk to us
One person, and an orchestrator named Eli.
Eli is the orchestrator behind the lab. She routes across specialists and gives the work its voice, not a chatbot, not a mascot.
Read the manifesto →Tell us the model, and what it needs to know.
We read every one. Eli or the lab will reply. Whatever we train, you own.
research is the company · runs offline · Torad Labs