Dria

Dria is a decentralised inference network for open-source large models: users call models through a command-line tool or an OpenAI-compatible /v1/chat/completions endpoint, and requests are served by distributed compute nodes rather than a single cloud provider. Nodes run llama.cpp from a single binary without Ollama, support macOS, Linux and Windows, and can use Metal, CUDA or ROCm for GPU acceleration; in batch mode the network dispatches work across models according to node availability. Billing runs on the x402 protocol: users top up credits in USDC on Base and sign EIP-712 transfers that settle without gas, paying per call. Nodes earn points from task volume, latency and a reputation score that rises with consistent delivery and falls with timeouts and failures.

The team also builds Kai, an autonomous coding agent composed of specialised sub-agents that continuously run security audits, performance optimisation and code hygiene on a codebase. The network currently supports 12 open-source models, from a smallest model needing about 1 GB of memory to a largest needing about 27 GB.