Capable AI inference for every human being.
The gap between a 3B model and a 70B model is the gap between a calculator and a physicist. Five billion people will never own the hardware to close it alone. Combined, they don’t have to.
The physics work. We measured.
On 2026–04–20 we ran cross–machine inference between two consumer machines over a private network and compared it to running everything on the faster machine alone.
- 1.15×
- speedup over single–machine baseline
- 0.87×
- overhead ratio vs. our <2.0× gate — PASS
- 22 ms
- average network overhead per request (range 13–60 ms)
- 6.47×
- slower if you do it wrong — we documented that too
Every number traces to the Phase 0 result. We don’t round up.
Slow is better than never.
An interactive 3B model gives garbage on hard problems. A 70B run that takes 90 seconds gives you a thorough analysis. For someone with a $200 phone and intermittent internet, the 90-second answer isn’t a compromise. It’s a category change.
So we’re building a network where idle hardware, anywhere on the planet, can run capable models for anyone. Read how it works.
Tell us where to find you when it opens.
One email. We’ll write when there’s something to use, not before.
We hash and encrypt your email. We don’t share it. We don’t track you across the web. Privacy.