Modes & Routing

Modes

Mode	Latency	Trust	Description
direct	<1s	Relay-managed	Centralized GPU providers
opengpu	1-2s	Trustless	OpenGPU Network - permissionless, redundant computing, on-chain settlement

OpenGPU task_address values are verifiable on ogpuscan.io.

Usage

{
  "model": "gpt-oss:120b",
  "messages": [...],
  "mode": "direct"
}

• auto (default): prefers direct, falls back to opengpu

• Not all models support both modes. See Models.