// a cooperative AI network made of gaming PCs

Your GPU
pays for your AI.

Lend your graphics card in the hours you don't game. In return, use AI models that are too big for your PC — split across many home GPUs, like one big machine.

Pre-register your rig See the launch models

one model · three home GPUs · one answer

160–180tokens/s from a 3-GPU chain of Qwen3.8 Flash Next, 16 conversations
< 2.5 sto the first word, with GPUs in two countries
0 / 222honest answers wrongly rejected by our verification; cheaters caught 128 / 128

Measured by us on 12 Nov 2026 on rented cloud GPUs in Romania and Czechia. Home networks will vary.

How it works

  1. 01

    Install the app

    A small open-source app runs on your PC. You choose when it works: only when you are not gaming, only at night, or whenever the PC is on anyway.

  2. 02

    Your GPU joins the network

    It answers requests on its own, or becomes one stage of a Chain: a large model split across several GPUs. Every answer is checked, so honest work counts and cheating doesn't.

  3. 03

    You earn access

    Verified work earns you credits. You spend them on models your card could never run alone: chat, coding, reading long PDFs.

The Chain

A big model doesn't fit on a gaming card. So we cut it into slices.

Each GPU holds a few dozen layers. A request flows through the cards in order, like an assembly line, and many conversations share the line at the same time. The full model never needs to sit on one machine.

We start with short chains of three or four cards in the same region, and lengthen them only when needed. 16 and 24 GB cards are the core; 12 GB cards fill the gaps.

  • modelQwen3.8 Flash Next
  • quantizationIQ3_XXS · 43.8 GiB
  • chain3 GPUs · 16 + 24 + 24 GB
  • throughput160–180 tok/s
  • vs one 48 GB card≈ 2.7×
  • networkRTT ≤ 25 ms · ≥ 100 Mbit/s

What you get — honestly

Access, not promises

Your work earns credits for models bigger than your own card. A cash payout comes only if and when the network sustains itself. We don't promise it.

Electricity is real

Your PC uses power while it works. We size credits so that the access you get is worth well more than the electricity you spend — at least three times its cost.

Offered hours ≠ paid hours

The network fills the most efficient cards first. Being available is rewarded a little; real verified work is rewarded most.

Launch models

Swipe through the roster we plan to launch with. Tell us what you want next.

BASE · FREE PLAN

Gemma 4 12B

Everyday chat and summaries. Runs on a single card.

to contribute
8 GB+
quantization
Q3

MEDIUM

Gemma 4 26B-A4B

Stronger reasoning, code and long documents. One 16 GB card, or a Chain of smaller ones.

to contribute
16 GB+
quantization
IQ4_XS

TOP · IN THE CHAIN

Qwen3.8 Flash Next

A frontier-class open model for coding, split across 3–4 home GPUs. LiveCodeBench v6: 86.3 in this quantization (publisher's figure).

to contribute
16–24 GB
quantization
IQ3_XXS

Subject to licence confirmation for hosted use.

NEXT?

Want a bigger model?

Another model, another quantization, a bigger one? Vote on the roster →

Open source where it matters

You run our code on your machine. You should be able to read it.

The host app and the Chain stage — everything you install — will be published as open source, together with the engine patches it builds on (a fork of llama.cpp). The repository opens before the launch.

$ publicompute-host --start --only-when-idle
› gpu: RTX 4070 Ti Super · 16 GB · ready
# source code published before launch

Shape the roster

Which models should the network run next? It takes 30 seconds.

What would you change?

Pre-register your rig

Self-declared, no install needed. It helps us know which PCs the network will have at launch — and which Chains we can build.

Not sure about your line? Chains need about 100 Mbit/s up and down and under 25 ms to the nearest servers. Our line test tool will be published with the app.

Questions

Will it slow down my games?

No: the app works only in the hours you allow, and stops when you start playing.

Can other people see what my GPU processes?

Your PC processes slices of a model's computation, not files of yours. We never access your files. Privacy options for users of the network are described at launch.

How do you stop cheaters?

A share of every node's work is recomputed by independent verifiers. In our tests honest nodes were never rejected and degraded models were always caught.

Is this crypto?

No. No coin, no speculation. Credits are an internal unit to use the network.

When do you launch?

When the network is ready and enough people have joined the waitlist. Pre-registered rigs are invited first.

Pre-register your rig