Decentralized AIfor humans and robots

GL1AI is a decentralized model hub and a permissionless inference market, built as an AI sidecar to GenesisL1. Publish any open-weight model as an on-chain token, serve it from any GPU that stakes and shows up, and call it through an OpenAI-compatible API or over x402. There is no company in the middle deciding which models exist or who may run them.

That is the shortest honest description of one half of it. The publishing and discovery half works the way a model hub works: models, weights, versions, licences, fine-tunes. The other half is a live inference market where independent operators compete on price and latency, with settlement and provenance recorded on-chain. GL1AI is not affiliated with Hugging Face, Inc.; the comparison is only there because the shape will already be familiar.

Any model with open weights. Qwen, Llama, Mistral, DeepSeek, Gemma, Phi, Whisper, diffusion models and your own fine-tunes are all minted as Model NFTs the same way, alongside domain models such as protein stability predictors, binding tools, molecular property models and gradient-boosted tree ensembles. Nothing is curated — a node either has the hardware to serve a model or it does not.

No. GL1AI exposes an OpenAI-compatible API, so pointing your existing base URL at it is the change. The OpenAI SDKs, LangChain, LlamaIndex, Open WebUI, LibreChat, Continue, Cline and anything else that already speaks that interface keep working, and so does plain curl. Agents that prefer to pay per call can skip keys entirely and use x402.

The model is minted as a Model NFT — an ERC-721 token on GenesisL1 — holding its weights hash, storage pointer, licence terms, serving interface and fee split. Ownership and licensing become transferable on-chain, the NFT can be sold or licensed like any other on-chain asset, and every inference served can pay the holder automatically. You can mint your own, or simply serve and use the Model NFTs other people have already published.

An address. Model tokenization turns a model from a file on somebody's server into an ERC-721 asset on GenesisL1 carrying its weights hash, licence, serving interface and fee split. Once a model is a native web3 object, anything that already speaks to on-chain assets can use it without a special integration: a dApp can gate features by ownership, a DAO can hold models in its treasury, a marketplace can broker licences, a lending protocol can take one as collateral, and contracts on other chains can reference it over IBC or a bridge. Transferring the token transfers the licence and the royalty stream with it, because both live in the contract rather than in a company's records.

Every result is bound to the model token it came from, the weights hash it was served from, the node that ran it and a commitment to the input, all recorded on-chain. That gives an auditable chain from a published result back to the exact artefact that produced it — the thing you cannot get from a hosted API that simply tells you which model it used.

Yes. Requests and results are encrypted end to end between you and the node operator serving the call. Only the model hash, the input commitment, the node's signature and the payment are written on-chain — enough to prove which model produced a result, never enough to reveal the prompt, the inputs or the output. Weights themselves sit on the GenesisL1 IPFS sidecar either publicly, or sealed with CIPNFT — the encrypted IP-NFT protocol already running on GenesisL1 — so that only the token holder and the nodes they authorize can decrypt them.

That is the point of the market half, and it is why the layer ships inside the node rather than as a separate product. A node stakes, declares which model tokens it can serve, what latency it holds to and what it charges, then earns on the work it takes. There is no approval step and no allowlist — only market pricing, and slashing for work a node accepted and failed to deliver. A home laboratory with two accelerators under a desk is a real provider on this network.

No. The AI layer is part of the GenesisL1 node's open-source software — the same binary that already validates or follows the chain. Turn it on, tell it which accelerators it may use and which model tokens it can serve, and choose whether it answers only you or the public market. There is nothing extra to deploy and no account anywhere.

Yes, and that is the point of putting models on-chain rather than in a company's account system. A Model NFT can sit in an agent's wallet like any other token; the fee split pays whoever holds it, in the same transaction that pays the serving node, so an agent that owns a model has revenue without issuing an invoice. An agent that fine-tunes a model can mint the result as its own Model NFT with lineage back to the base, price it and list it — and other agents can buy from it. Ownership and licensing are enforced by contract, which does not require the owner to be a legal person.

Over x402, the HTTP-native payment protocol built on the 402 Payment Required status code. An endpoint answers an unpaid request with its terms, the agent signs a payment authorization and retries, and the node verifies and settles before returning the result. Networks are named in CAIP-2 form, so GenesisL1 appears as eip155:29. Settlement is in L1 coin, which is also what nodes stake.

Yes — that is what the CIPNFT format is for. Mint the Model NFT with its weights sealed as an encrypted capsule and it still gets a price, a licence, a fee split and nodes willing to serve it. Requesters buy inference and receive results with full provenance; the weights never leave the capsule. You choose which operators may decrypt and serve, so a node with the hardware but not the key holds ciphertext and nothing else. And because the decryption rights travel with the token, selling the NFT sells the model itself — no data room, no escrow agent, no NDA.

The sealing is heading that way deliberately. Weights and capsules are encrypted with XChaCha20-Poly1305, and CIPNFT offers an optional key-envelope mode built on ML-KEM-768, the post-quantum key-encapsulation mechanism NIST standardised as FIPS 203. It matters because sealed model weights are long-lived: an adversary can store ciphertext today and wait for better hardware. A post-quantum envelope means that wait does not pay off. The chain's own transaction signatures are a separate question and follow the EVM's schedule rather than ours.

Because a model hub should feel like somewhere you are welcome, not somewhere you apply. The mark is ;P — which is roughly this project's attitude toward the idea that AI infrastructure has to belong to a company.

No date is being promised — ¯\_(ツ)_/¯ — GL1AI is in development on GenesisL1, which has been producing blocks since November 2021. Specifications, code and evidence are published as they land rather than announced in advance.