AI

A privacy-first multi-model AI inference network. 5 models live now, 55 across every major provider via OpenRouter — coming soon. Inference is memory-only, wiped on response, with a signed cryptographic receipt on every reply.

5 live
50 in build

Model Lane

A small router (7–14B) classifies intent in under 3 seconds and selects the right model. In Broadcast and Diffuse modes all live models fire at once.

ModelByTypeStatus
Claude Haiku 4.5Anthropictextlive
Claude Sonnet 4.6Anthropictextlive
Llama 3.3 70BMetatextlive
Kimi K2.7 CodeMoonshotAItextlive
Kimi K2.6MoonshotAItextlive
Claude Fable 5Anthropictextin build
Claude Opus 4.8Anthropictextin build
Claude Opus 4.8 FastAnthropictextin build
Claude Opus 4.7Anthropictextin build
Claude Opus 4.7 FastAnthropictextin build
Claude Opus 4.6Anthropictextin build
Claude Opus 4.6 FastAnthropictextin build
Claude Opus 4.5 SonnetAnthropictextin build
Llama 3.1 405BMetatextin build
GPT-5.5 ProOpenAItextin build
GPT-5.5OpenAItextin build
GPT-5.4 Image 2OpenAIimagein build
GPT Chat LatestOpenAItextin build
Gemini 3.5 FlashGoogletextin build
Gemini 3.1 Flash LiteGoogletextin build
Gemini 3.1 Flash ImageGoogleimagein build
Gemini 3 Pro ImageGoogleimagein build
Grok 4.3xAItextin build
Grok Build 0.1xAItextin build
DeepSeek V4 ProDeepSeektextin build
DeepSeek V4 FlashDeepSeektextin build
Qwen3.7 MaxQwentextin build
Qwen3.7 PlusQwentextin build
Qwen3.6 Max PreviewQwentextin build
Qwen3.6 FlashQwentextin build
Qwen3.6 35B A3BQwentextin build
Qwen3.6 27BQwentextin build
Qwen3.5 PlusQwentextin build
Mistral Medium 3.5Mistral AItextin build
Nemotron 3 UltraNVIDIAtextin build
Nemotron 3 Nano OmniNVIDIAtextin build
North Mini CodeCoheretextin build
Granite 4.1 8BIBMtextin build
MiniMax M3MiniMaxtextin build
GLM 5.2Z.aitextin build
GLM 5.1Z.aitextin build
Ring-2.6-1TinclusionAItextin build
Ling-2.6-1TinclusionAItextin build
Ling-2.6 FlashinclusionAItextin build
Hy3 PreviewTencenttextin build
MiMo-V2.5 ProXiaomitextin build
MiMo-V2.5Xiaomitextin build
Laguna XS.2Poolsidetextin build
Laguna M.1Poolsidetextin build
Step 3.7 FlashStepFuntextin build
Perceptron Mk1Perceptrontextin build
Nex-N2-ProNex AGItextin build
FusionOpenRoutertextin build
Owl AlphaOpenRoutertextin build
Pareto Code RouterOpenRoutertextin build

Architecture

Inference pipeline: Router → Retrieval (optional) → Inference → Stream + Receipt → Wipe. Nothing is persisted.

Protocol
Rust, actix-web, WebSocket, SSE
Inference proxy
Stateless Rust binary, hash-published
Inference
vLLM / SGLang, open-weight only
GPU
RTX PRO 6000 Blackwell Max-Q, 96 GB ECC
Edge
Cloudflare Workers, IP stripping
Storage
None — memory-only
Receipts
Ed25519, offline key, published pubkey
Telemetry
None

Cryptographic Verification

Every response commits to its production conditions in a signed io_receipt_v1. Verifiable offline against the published Ed25519 public key — via SDK, MCP, CLI, or web.

receipt_v1:
model: claude-opus-4.8
policy: io-boundary-0.3
worker_binary_hash: sha256:9f1e...c2a4
enclave_quote: tee_attestation_v1
retained_prompt: false
signature: 0x8a3f...

Today — Inspectable

Inspect source, watch the network tab, confirm no client-side storage, verify the signature binds response to model and policy. Misattribution and client-side leakage are detectable.

v1.0 — Hardware Attested

Workers run inside TEE-attested enclaves. The enclave_quote field binds the receipt to a measured binary in hardware-isolated environment. retained_prompt: false becomes hardware-backed, not operational.

$IO Token

Fair launch on PumpFun (Solana) — no pre-sale, no venture allocation. $IO coordinates access and payment across the network.

Payment

Inference paid in $IO via x402 micropayment protocol — discounted vs other assets.

Access

Premium features — higher rate limits, broadcast/diffuse, proof certificates — gated by $IO.

Burn

Buyback-and-burn from net protocol revenue (after GPU & bandwidth) — supply tightens with usage.

Governance

At v1.0, holders vote on model registry, redaction policy, fee splits, and warrant canary cadence.

Fee policy note: Buyback-and-burn operates on net protocol revenue after infrastructure costs (GPU + bandwidth), not gross inference fees. The precise split between burn and operations reserve will be published as an on-chain parameter before activation.