Hugging Face is v1’s first endpoint vendor. The marketplace is designed around provider commitments and tariffs so later Gateway adapters can support additional inference vendors and GPU networks.
Providers can follow demand
Providers do not need to bootstrap every model speculatively. They can evaluate trading activity and market traction, decide which unserved demand is worth the cost of deployment, and register an offer when they are ready to serve it. Once the first offer becomes active, the existing model credit becomes the inference payment asset. Additional providers can enter later and compete through independent tariffs and service quality.Provider offer
An on-chain offer stores:- Model ID.
- Operator and payout addresses.
- Endpoint manifest commitment.
- Capability commitment.
- Metadata commitment.
- Current and pending tariff versions.
- Creation time and status.
Tariff
Each provider independently sets:| Charge | Calculation |
|---|---|
| Input | input tokens × input credits per million / 1,000,000 |
| Output | output tokens × output credits per million / 1,000,000 |
| Minimum | Lower bound on the usage charge. |
| Wake | Added when the endpoint must wake from a sleeping state. |
Routing
For automatic selection, the Gateway evaluates active verified offers and chooses the lowest maximum credit charge. Ties are broken deterministically by offer ID. Before routing, the Gateway checks:- On-chain offer status and commitments.
- Verification freshness.
- Endpoint health and wake state.
- Model context length.
- Request capability requirements.
- Permit balance and optional offer pinning.