seed
Serve layers while you sleep.
Your machine holds a slice of a model, say layers 24–35 of a 70B, and serves it whenever you're not using the GPU. Like seeding a file, minus the ratio anxiety.
open protocol · in development
Open models need a BitTorrent.
podtorrent pools idle compute from people's own machines into one swarm. 4090s, Mac Studios, Strix Halo boxes, homelab racks, data centres. Work gets split across peers. Only verified work earns.
Got a GPU? It could be here.
— peers
— orbital
— PFLOPS
distances not to scale · altitudes compressed · off-world nodes simulated
simulated network — protocol in development
01 · the napster moment
The weights are free. The compute isn't.
Llama, Qwen, DeepSeek, Mistral, Gemma, gpt-oss. Anyone can download them. That fight is won.
Running the big ones is another story. A 70B model at 8-bit wants about 70 GB of fast memory. A 405B wants about 400. Almost all of that lives in a few data centres, owned by a few companies, rented back to you by the token.
Open weights on closed iron is half an open ecosystem.
02 · seed, leech, swarm
BitTorrent, but the pieces are tensors.
Same idea that moved the internet's files, pointed at compute instead.
leech
Run models bigger than your box.
Borrow the layers you can't fit. A laptop drives a 70B by streaming activations through peers who have them loaded. Credits you earned seeding pay for it.
swarm
No single anything.
The scheduler splits each job by VRAM, bandwidth and latency. A slow or vanished peer gets routed around. More peers, more paths, more headroom.
03 · proof of useful work
Only verified work earns.
Every shard gets spot-checked. A random peer re-runs a sample. Cheap checks run on the activations. Results that don't match get rejected, and the peer that sent them gets ignored.
Got a 40-GPU farm in a shed? Bring it. Scale is fine. Faking it isn't.
- 0x3fa1… layers 0–11 · Lagos · RTX 3090 verified ✓
- 0x91c4… layers 12–23 · LEO-7 · orbital (sim) verified ✓
- 0x77e0… layers 24–35 · Chennai · M2 Max verified ✓
- 0x0bad… layers 36–47 · ??? · "RTX 9090" mismatch · rejected ✗
- 0xc0de… layers 48–59 · Reykjavík · 2× 7900 XTX verified ✓
04 · the payback meter
Idle hours in, credits out.
Pick your hardware and how long it sits idle. See roughly what that buys you on the swarm.
credits / month illustrative, not a promise
≈ — hours of a 70B model
Credits today. Verified compute has value — we're building the rails for that too.
05 · bedrooms → data centres → orbit
Wherever there's compute, there's a peer.
- rigs merge into one mesh around the planet
- the orbital shell lights up
- Earth–Moon, and a relay line heading for Mars
Orbital and off-world nodes are simulated. The protocol doesn't care where the silicon is. The scheduler cares how far away it is.
06 · early inference providers
The first nodes light up first. Claim yours.
We're onboarding the first inference providers by hand. Tell us what you've got. We'll be in touch when your hardware class is up.
github · community & spec — coming soon