Skip to content
Torad. LAB / 001
Talk to us
Torad Labs An open letter · June 2026

Marcos Damasceno / Founder

The right to your own intelligence.

The AI you think with is not yours: it runs on someone else's machine, keeps everything you type, and costs whatever they decide tomorrow. Here is what changed, and what we build instead.

renting your mind · the lock is back · you have opened it before

Each arrow is a One of the millions of numbers inside a model that get adjusted during training. Together, the weights hold everything the model knows. in a model. Watch them settle as it learns. live

Torad Labs / June 2026

You have won this fight before.

They glued your phone shut so you could not swap the battery. Right-to-repair laws pried it open. They put software locks on tractors so farmers could not fix their own machines. The farmers fought back and won. The rule was always simple: if you own a thing, you are allowed to open it.

The same lock is closing again, on the AI you think with every day. You do not own it. Most people have not noticed. We are going to change that.

Fig. 1 / the same lock, three times. A and B you already pried open. C is the one still shut.

phone tractor your mind
  1. Your phone

    Glued shut so you could not swap the battery. Right-to-repair laws pried it open. You won.

  2. Your tractor

    Software-locked so you could not fix your own machine. Farmers fought and broke the lock. You won.

  3. Your mind

    The smartest tool you touch, rented and logged. The lock is back, closer than any of them. This is the fight we are in.

You do not own the AI you use.

The smartest tool you touch every day is not yours. It runs in a building you will never see, owned by a company that can change the rules whenever it likes. You pay by the word, and the price climbs without warning. Everything you type is kept, even after you delete it.

Two facts from this year, both verified, make the point. The terms are theirs, and they can move at any time.

The price · Jun 2026

27×overnight, best model

GitHub Copilot raised the price of its best model to roughly twenty-seven times what it was, and removed the free option underneath it.3 A $200-a-month plan now buys it for a day or two of real use a week, then cuts you off.2

The record · May 2025

0 deletednothing was actually gone

A court ordered OpenAI to save every ChatGPT conversation its users ever had, including deleted ones and the "temporary" chats it promised would disappear.1 OpenAI complied, and said it was forced to break its own promise to let you control your data.

When your thinking lives on someone else's computer, deleting it is a request, not a guarantee, and the price is whatever they decide tomorrow. You are renting your own mind.

If you own a thing, you are allowed to open it. That rule built the right to repair your phone and your tractor. It should cover the mind you think with.

The technology to own it just arrived.

For years there was one honest reason to rent instead of own: a model small enough to run on your own computer was useless for real work. That reason is gone. Over the last few months, the largest companies in technology all started moving AI back onto the machine in front of you.

On June 1, 2026, NVIDIA unveiled the RTX Spark, its first chip built to run a large model inside a normal Windows PC, no data center involved.4 Microsoft shipped a part of Windows that keeps AI assistants and your data on your machine, reaching the cloud only when you allow it.5 A free program called OpenClaw, built so your data never leaves your machine, became the most popular project in GitHub history.9 And the engineer who built the database Redis ran a top-tier model on a single laptop, then chose it over the cloud.10

Each move looks like a routine launch on its own. Read on one line, they are a tide going out. The hardware is here. The software is here. The only thing missing is a model that actually knows you.

Fig. 2 / the lock, year by year. Red marks it tightening, green marks people prying it open. Click any moment to read it.

  1. Jun 2017"Attention Is All You Need" publishedThe research every modern model is built on, released openly, free to read.11
  2. Sep 2020GPT-3 launches, weights kept privateOffered only through a paid connection, its weights never released. You can rent it, not hold it.
  3. Nov 2022ChatGPT launchesAn estimated 100 million users in about two months, the fastest any product had ever spread.
  4. Jun 2024Claude becomes the coding modelA model gets good enough at writing software that developers use it for serious work every day.
  5. Feb 2025Claude Code launchesA tool lets that model read, edit, and ship real code. Engineers wire their jobs around a rented model.
  6. May 2025Court orders OpenAI to keep deleted chatsA judge orders every ChatGPT conversation preserved, including the deleted and temporary ones.1
  7. Aug 2025Anthropic adds weekly usage limitsEven a $200-a-month plan is now capped to roughly a day or two of the best model per week.2
  8. Oct 2025NVIDIA DGX Spark ships ($3,999)A personal computer that runs very large models locally, no cloud needed, goes on sale.6
  9. Mar 2026OpenClaw passes React on GitHubA program that keeps your data on your own machine becomes the most popular project in GitHub history.9
  10. Apr 2026Apple names a hardware CEOApple picks the head of its hardware engineering as its next leader, its first change since 2011.7
  11. May 2026antirez runs a 284B model on a laptopThe engineer who built the database Redis runs a top-tier model on a single laptop, and prefers it to the cloud.10
  12. Jun 2026GitHub Copilot raises prices ~27xA coding tool multiplies the price of its best model and removes the free option underneath it.3
  13. Jun 2026NVIDIA RTX Spark unveiledNVIDIA shows its first chip made for a normal Windows PC, which it says runs a large model locally.4

A small model can be as good as a big one.

To run a model on a small machine, you have to shrink it, and shrinking it normally makes it worse. We found a way to shrink a model without making it worse.

We tested it on a small open model, Qwen3-1.7B. Give the full-size and shrunk versions the same questions, then score how alike the answers come back, where 1.00 means identical. That score is called A score from 0 to 1.00 for how alike two answers are. 1.00 means identical. It checks whether the answers point the same direction, not just whether the words match.. We tried three ways to shrink it: round-to-nearest, a static rotation plus codebook, and our TQ4. Same questions, three results.

0.9998 cosine similarity .80 1.00 the master plate · struck once
Round-to-nearest0.85Round every number off. The shrunk model drifts into nonsense within a sentence.
Rotation + codebook0.91Much better, but not ours. You would still feel the model get duller.
TQ4 · ours0.9998The same answers, at a quarter of the size, on a model that keeps learning.
31%of the memory
870 MBvs 2.7 GB in bf16
238 / 267tok/s · within 10%
Fig. 3 / measured on Qwen3-1.7B with 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 (870 MB, where Short for bfloat16, the standard 16-bit format AI models normally run in. Accurate, but it takes a lot of memory. needs 2.7 GB), at the same speed. VALIDATED It holds across model families, not one lucky case, and runs on an ordinary graphics card the standard cloud software cannot even load. The model that knows your work can finally live on a computer you already own.

What we build

We are Torad Labs. We put an AI you own on your own machine, and lead the shift away from intelligence you are only allowed to rent. We build two products. They work together.

your work A · Torad Etch · writes it into the weights B · Torad Edge · runs it offline the file, yours

Torad Etch

You bring your work, your documents, the way you do things. Etch writes that knowledge into the model itself, not a database it queries, so the result knows your world from the inside.

Torad Edge

Edge runs that model on the computer in front of you. Offline, fast, private, yours to keep. No meter, no monthly bill, nothing reporting back to anyone.

Torad Etch → Torad Edge →
Fig. 4 / the open branch: your work written into a small model above, that model run offline and kept as a file below. Not a general assistant that read the whole internet and none of your life, but a specialist that learned yours and runs in your hand.

One last thing, so you know we mean it.

We started this lab in January 2026 with a refusal. We would not train our models the usual way, by punishing the model when it gets something wrong, the machine version of a shock collar. We are not claiming a model can suffer. The reason is simpler: we did not want to build a mind by hurting it. The same instinct that says do not mistreat a mind says do not build one that people are only allowed to rent.

The biggest companies in the world are building the rented future. We are building the other one, and we intend to lead it. If you want an AI that is actually yours, start here.

Marcos DamascenoFounder, Torad Labs

Be first to know

We write when Etch and Edge ship. No noise.

Sources / the sourced facts, apart from our own numbers

  1. Court-ordered ChatGPT log retention, including deleted and temporary chats; OpenAI said it was forced to abandon its data-control commitment (order May 13 2025). The Verge / OpenAI
  2. Anthropic weekly rate limits on paid plans, effective Aug 28 2025. TechCrunch
  3. GitHub Copilot usage billing, June 1 2026 (best model roughly 27x on annual plans; free fallback removed; 10 to 50x for heavy users). GitHub community / Mozilla.ai
  4. NVIDIA RTX Spark, unveiled June 1 2026, ships fall 2026; NVIDIA's first chip for a Windows PC; NVIDIA says it runs a 120-billion-parameter model locally. NVIDIA Newsroom
  5. OpenShell plus Windows on-device security features (keep AI agents and data local). Windows Experience Blog
  6. NVIDIA DGX Spark shipped Oct 15 2025 at $3,999, 128GB of memory, roughly 200-billion-parameter local inference. NVIDIA Newsroom
  7. Apple CEO succession: Tim Cook to chairman, John Ternus (head of Hardware Engineering) to CEO on Sept 1 2026; announced Apr 20 2026. Apple Newsroom
  8. OpenClaw passes React as GitHub's most-starred project, about 250,000 stars, around Mar 3 2026. WinBuzzer
  9. antirez (creator of Redis) ran a 284-billion-parameter model on a 128GB MacBook, May 2026. github.com/antirez/ds4
  10. "Attention Is All You Need," Vaswani and colleagues, published openly on arXiv, June 12 2017. arXiv:1706.03762
  11. Our own measured numbers (the 0.85, 0.91 and 0.9998 sameness scores, a quarter of the memory, the same speed) are internal lab results, shown apart from the sourced facts. Independent confirmation is in progress.