The commons
Compute that scales with humanity.
A cooperative network where anyone can contribute idle hardware and access capable models in return. The network never sleeps because the world never sleeps.
Four modes of shared inference
One platform, four mechanisms. Each fits a different latency profile. The user never picks; the platform picks the best available mode for the request.
Local solo
The model runs on your device. Zero network overhead, perfect privacy. Limited by what your hardware can hold. The fast path when it’s available.
LAN pooling
Several machines on the same physical network combine memory to run a model that none could run alone. Each holds a slice of layers; data flows machine–to–machine over a switch. Libraries, schools, co–working spaces, neighborhood co–ops — any group with proximity and a few machines.
Speculative decoding
Your device runs a small draft model to propose tokens fast; a remote machine runs the large model to verify. Accepted tokens stream immediately. The person with a $300 laptop gets interactive 70B inference through a contributor’s 128GB machine across the internet.
Async batch
Submit a prompt, get a job ID, receive the result when it’s done. Fifteen seconds to two minutes is normal. For document analysis, code generation, summarization, translation — the majority of high–value AI tasks — this is plenty. A 90–second wait for a thorough 70B answer is infinitely better than instant garbage from a 3B.
Credits, not currency
Contributors earn credits for the compute they provide. Users spend credits for the compute they consume. The exchange rate is a simple accounting parameter, not a market. No tokens. No speculation. No volatility.
A small device contributing for hours can earn enough credits to run a large model occasionally. That’s the goal: the smallest contributor still has access.
What we are not doing
- Not a cryptocurrency. Credits don’t leave the platform as a tradeable asset.
- Not a corporate cloud. We’re building community infrastructure that happens to serve enterprise too.
- Not a closed ecosystem. The serving engines are open source; the API is open; the models are public weights.
- Not a promise we haven’t verified. We just published the Phase 0 result. The next phases will be public the same way.
Why now
The gap between what a frontier model can do and what a small model can do is widening, not closing. If access tracks hardware ownership, AI becomes the largest inequality multiplier in modern history. If access tracks willingness to share idle compute, it becomes a public good.
The technology to do this exists today. The economics work today. The community pieces — libraries, schools, maker spaces — have been waiting for someone to wire it up.