IRIS PROJECT SUMMARY

Current Status

1. Token IRIS live on Base · contract address pending publication.
2. Vision-mesh alpha: 4 cold-running nodes across EU + NA, serving inference for a private alpha cohort.
3. First open model (iris-vision-70b) targeted for HuggingFace publication this quarter.
4. iOS companion app in TestFlight with the alpha group — camera streaming, share-sheet ingestion, persistent visual memory.
5. OpenAI-compatible vision API already live — drop-in replacement for any builder using gpt-4o's image_url messages.

Summary

IRIS is a multimodal AI agent network on Base. The premise is simple: your phone has a camera. Your agent doesn't. IRIS closes that gap with open-source vision-language models served by a peer-to-peer mesh, billed in $IRIS.

The product is two layers. On top, a mobile app that funnels camera input, screenshots, screen-records, and file drops into a persistent agent session. Underneath, an open inference network that anyone with a GPU can join.

Why IRIS, why now

Three forces converging:

1. Vision is the next API contract. Every consumer AI app shipping in 2026 has a camera button. Builders are paying premium pricing to closed labs whose vision models refuse half the queries that matter (medical, legal, NSFW, financial-doc, etc.).
2. Consumer GPUs idle 22h/day. Hundreds of millions of cards globally, mostly running screensavers. Aggregate compute dwarfs every hyperscaler combined. Routing inference to them is the obvious arbitrage.
3. Mobile compute can't catch up. Even with Apple Intelligence and Gemini Nano, the heaviest vision-reasoning workloads still need a real GPU. Sending the photo to the network and getting an answer back beats waiting for local inference on a phone.

Models

The IRIS model line is vision-first. We start from open frontier checkpoints (Llama 3.3, Mistral 24B, Gemma 9B) and apply a custom visual-instruction-tuning toolchain on a dataset weighted heavily toward screenshots, document layouts, photos, and screen-records.

• iris-vision-70b — flagship. Llama 3.3 base. 4× A100 or 1× H100.
• iris-vision-24b — workhorse. Mistral base. Single 4090.
• iris-mobile-9b — lightweight. Gemma base. Runs on edge devices including high-end phones.

All weights publish to HuggingFace under iris-network/. We don't gate access — the Iris Network is just the cheapest place to run them.

Network — Peer-to-peer Vision Compute

The IRIS Network is a peer-to-peer mesh. Nodes run a small client that advertises capacity (GPU, models supported, thermal headroom) to a gossip routing layer. Requests resolve to the lowest-latency idle node. Every served token generates a signed attestation that batches onchain to the rewards contract per epoch.

• Pool routing — requests find the best-matching idle node.
• Reputation weighting — repeat divergence from consensus cuts node payouts.
• Onchain settlement — every epoch finalizes payouts on Base.
• No staking — joining is free, leaving is free.

Architecture

1. Edge nodes — GPU operators running the IRIS client.
2. Routing layer — peer-to-peer overlay matching requests to nodes.
3. Settlement layer — onchain on Base, batched per epoch.
4. Mobile + API gateway — thin proxy at api.eyeofiris.tech and the iOS companion.

Security model

• Output divergence detection — N=3 redundancy on a sample of requests. Repeat divergence cuts reputation.
• No persistent prompt storage — image data is held only for the duration of a single inference call.
• Sybil resistance — hardware attestation at node registration. Duplicate hardware claims fold into one rewards entity.
• Optional TEE mode — sensitive workloads can opt into a TEE-backed node pool (beta).

Roadmap

1. Q3 — Alpha network opens. iris-vision-70b drops on HuggingFace. Mobile app TestFlight expands.
2. Q4 — Onchain payouts go live. $IRIS market opens on Base.
3. Q1 — Visual memory layer (private vector index per user). Multi-tenant on-node isolation.
4. Q2 — iris-vision-24b drops. Vision API beta exits beta. Android companion ships.
5. Q3+1 — Distributed fine-tuning MVP (community-trained vision adapters).

Economics & $IRIS

$IRIS is the unit of account inside the network. Apps pay for inference in $IRIS (auto-converted USDC at the gateway). Node operators receive $IRIS. The protocol takes a 1% fee on inference revenue, converts to $IRIS, and burns — creating deflationary pressure tied to network usage.

• Total supply: 500,000,000 IRIS (fixed).
• 40% — node operator rewards pool, distributed over 4 years.
• 30% — initial LP on Base (Uniswap V4, 0.3% fee tier).
• 15% — team & contributors, 4-year linear vest, 1-year cliff.
• 10% — ecosystem grants & integrations.
• 5% — public airdrop to alpha-period operators and mobile beta users.

Open source

The client, the routing layer, and the settlement contracts are all Apache 2.0. Model weights publish under their respective base-model licenses. Every artifact is reproducible by any contributor — no proprietary fork of any base model.

FAQ

Why Base instead of another L2? Cheap settlement, strongest onchain identity tooling, and the most active retail flow of any L2. Plus Coinbase as a default custodian for the long tail of users.

Is this just another decentralized compute network? Vision is the wedge. Most existing decentralized inference projects ship text-only and treat vision as a future quarter. We're vision-first from the model line down through the mobile app.

What about regulation? Open weights, decentralized routing, software-mediated payouts. Operators own their own jurisdiction. We don't host or moderate content.

Can I self-host the gateway? Yes — the gateway is a thin proxy. One binary plus a config file. Documented in Docs.

What's special about the rainbow? Iris is the Greek messenger goddess whose path was the rainbow connecting Olympus to mortal earth. Our brand mark — an eye with a rainbow iris — is the literal expression of what IRIS does: the connection between vision and intelligence, drawn as light passing through the eye.