IRIS runs on a peer-to-peer mesh of consumer GPUs that serve vision inference around the clock. Anyone can plug in a card, anyone can call the API. Settlement and reputation are onchain.
Got an idle 3090 / 4090 / A100? Install the IRIS client, register your wallet, earn $IRIS per token served. No staking, no penalty for stopping.
READ THE DOCS →Vision-capable models with no per-minute rate limits, no content guardrails, ~30% cheaper than centralized vision APIs. Drop-in for any builder.
API REFERENCE →Requests route to the lowest-latency node that can serve the requested model. Camera streams stay local, document inference goes wherever the spot price is best.
Every epoch, node operators are paid in $IRIS proportional to served tokens, weighted by a peer-verification quality score. Payouts settle on Base.
Prompts and image data never persist on serving nodes beyond the inference call. Optional TEE-backed mode for sensitive workloads in beta.
Open weights, open client, open settlement contracts. Leave whenever. Drain your $IRIS, uninstall the client, take your reputation with you.