Run models too big for your machine.
Warren pools the idle memory sitting in other people's laptops and desktops into one shared swarm — and pays you for the slice you contribute.
Warren pools idle memory from thousands of devices into one shared swarm, so you can run models that would otherwise need a rack of GPUs you don't own.
a 70B model, split into six shards, run in parallel
The GPU you don't have
A capable model needs more memory than one consumer machine holds. Cloud instances fix that, at cloud prices, billed whether you're using every cycle or not.
One laptop, maxed out
Warren, six peers
How a swarm actually runs
Four steps, most of them automatic.
Register
Install the agent. It benchmarks your free RAM, VRAM, and bandwidth, and reports only that — nothing on disk is read.
Match
You're grouped into a swarm of peers with compatible specs, sized to the model a runner wants to use.
Compute
The model is split into shards. Each device holds and runs one slice; partial results pass between peers over an encrypted link.
Settle
Credits move automatically at the end of the session — runners spend per compute-second, contributors get paid for the same second.
One endpoint, hundreds of devices behind it
Every node below stands in for real hardware sitting in someone's home. Hover or tab through one to trace its path through the swarm to the single endpoint your app talks to.
Two sides of one swarm
Every device in the network is either lending capacity or borrowing it. Most people will do both.
Contribute compute
Your device sits idle most of the day. Let it join a swarm and earn credits for the memory and cycles it isn't using.
- Minimum free RAM8 GB
- Runs whileplugged in, idle
- Payoutper compute-second
- Your filesnever exposed
Run a model
Point Warren at a model larger than your own hardware, and it's sharded across a matched swarm on demand.
- Supported sizesup to 200B params
- Billingpay-as-you-go credits
- Cold startunder 40s typical
- APIOpenAI-compatible
Before you ask
The agent runs the model shard in an isolated process with no access to your files, browser data, or other applications. It only reads the memory and compute cycles you've marked as shareable.
No. Runners talk to the swarm as a single endpoint. Device identity and location stay inside the matching layer.
Credits accrue per compute-second and settle at the end of each session. You'll be able to hold them or cash out once payouts open.
We're starting with a small set of open-weight models in the 30B–200B range, chosen for how well they shard across heterogeneous hardware.
Be in the first swarm.
We're onboarding contributors and runners in small batches ahead of public launch. Tell us which side you're on.