Product
Call it like any other model.Serve it, and you are the one being paid.
One network seen from two ends — an endpoint for the application calling it, an open validator set for the hardware answering it. The token that pays the second is what keeps the first free of a vendor.
Renting a mind has three failure modes.None of them are about capability.
The frontier models work. What does not work is the arrangement you have to accept in order to call them.
The endpoint is a tenancy
The moment it becomes private property, humanity becomes a tenant in its own mind.
The price, the terms and the model behind the name can all change under you, and a tenant has no standing to object. Nothing you built on top of it is yours in the way you assumed.
The meter never stops
You will not pay for tokens by the hour. You will hold them.
Usage pricing puts a floor under every call you will ever make, and no amount of volume removes it. The cost of thinking is a rent, paid forever, to someone else.
The claim cannot be checked
We reject the idea that safety requires secrecy.
Closed weights make every statement about a model's behaviour a statement you have to take on trust. Not because it is false, but because there is no way for you to find out.
Three mechanisms.None of them require trusting us.
Every one of these is defined in the codex and can be checked against it. None of them describe an intention.
An open validator set
Anyone whose hardware clears a tier's floor can register and be routed to.
The routing service selects from the eligible pool for the surface and latency tier requested. There is no allowlist to be admitted to and no seat to be granted.
Sampled audit, and a challenge
A share of queries is re-run on a second validator and the two outputs compared.
A mismatch opens a challenge through the challenge service. Serving dishonestly stops paying long before it becomes worth anyone's while to detect.
A treasury the network refills
Validators are paid from the swap fees on the token's own liquidity, not from new issuance.
Supply is fixed at 1,000,000,000 OGI at genesis and the mint authority is revoked. The network runs on what it earns, which is why it does not need a funding round to keep running.
Every swap funds a small slice of the network's continued operation.The cycle is the network's metabolism.
The treasury seeds and holds the liquidity position itself, so the fees it earns while the token is traded are what pays for compute. The token contract carries no transfer tax; only the protocol's own position is in the cycle.
Validators are not paid out of new issuance, because there is none. They are paid out of what the token's own liquidity earns while it is traded — which is why the network does not need a funding round to keep running.
Seven phases.The last one removes the operator.
The network is in Phase 1, language surface. A phase ends when its exit criteria are met, not when a date arrives.
Phase 0
complete
Genesis
- Treasury revenue at or above operations baseline, sustained
Phase 1
current
Language surface
- Sustained inference throughput
- Audit divergence below the published threshold of 0.0010
- At least 250 independent validators across at least two jurisdictions
Phase 2
planned
Vision and multimodal
- Cross-modal queries served with documented latency and quality
Phase 3
planned
Reasoning and memory
- Multi-step plans complete above the published verified-correctness rate
- Memory recall verified by sampled re-execution
Phase 4
planned
Embodiment
- Published count of registered embodiments reached
- Embodied episodes flowing into the training corpus at a sustained rate
Phase 5
planned
Continual learning
- Consecutive checkpoint promotions passing the forgetting eval on all historical surfaces
Phase 6
planned
Operator sunset
- No foundation-distinguished role remains active
Four tiers.Serve with the hardware you already have.
A tier is a hardware class, not a price plan. It decides which surfaces you are eligible to serve and which roles you can take on.
| Tier | Reference hardware | Roles | Minimum stake | Latency SLO |
|---|---|---|---|---|
| Light | Single consumer GPU, 24GB+ | inference:distilled evaluation storage audit | 25,000 OGI | 2,400 ms |
| Standard | Single consumer/prosumer GPU, 32GB+ | inference evaluation audit | 60,000 OGI | 1,600 ms |
| Heavy | Single data-center GPU, 80GB | inference evaluation training:worker audit | 200,000 OGI | 900 ms |
| Cluster | Multi-node, high-bandwidth interconnect | inference:frontier training:full evaluation | 750,000 OGI | 700 ms |
Missing the latency SLO costs a validator its place on the routing table until it recovers. It does not cost the stake.
Three ways in.Call it, serve it, or hold it.
Call it
Point any OpenAI-compatible client at the gateway — the same one you already use for Kimi or Qwen. Phase 1 runs a free public endpoint, so there is no key to request first.
Read the APIServe it
Register in the tier your hardware clears, take the roles it allows, and get routed queries. You are paid out of the treasury, per attested job.
Validator specHold it
Every token is a GPU hour. The fees that refill the treasury come from the token's own liquidity, so holding it is how the network is funded.
Token chapter