M-NFT (Model-as-NFT)
🧠 What is M-NFT?
M-NFT is ASTRAI’s core asset abstraction: every trained AI model is encapsulated as a tokenized, programmable, composable NFT, making it:
Ownable → Creators, fine-tuners, and DAO backers are recorded as co-owners
Monetizable → Any usage generates revenue that flows to stakeholders
Composable → Can be used across dApps, protocols, or nested in agent structures
Programmable → Can enforce access control, royalty rules, retraining logic, governance, etc.
📦 M-NFT Structure
Token Layer
ERC-721 or ERC-6551 token (with smart contract extensions)
Metadata Layer
Includes:
Model architecture (e.g., LLaMA-2, Whisper)
Checkpoint hash
Training dataset references
Contributor list + weights
Inference endpoint info
Access rights
Storage Layer
Weights stored on Arweave/IPFS/Filecoin
Signed training logs + dataset hashes stored on-chain or verifiable data layer
Logic Layer
Smart contracts define:
Usage permissions (open, allowlist, pay-per-call)
Royalty splits
Retraining/upgrading protocol
Governance & transfer conditions
🔐 Usage & Licensing Framework
M-NFT enforces on-chain access control and usage rights, allowing for:
Open Access: Anyone can call the model via endpoint
Token-Gated Access: Only holders of specific tokens/NFTs (e.g. DAO members) can use
Pay-per-Call: Usage metered via off-chain logs + signed receipts (oracles), payment streamed to contributors
Subscription Licensing: Optional ERC-4907 style lease contracts for time-based model rental
All calls are:
Logged and hashed by gateway (e.g., API Gateway or Rollup-based zk proof system)
Verified against M-NFT logic
Billed and routed via smart contracts
⚙️ Key Tech Integrations
Ownership & Revenue Splits
ERC-721 / ERC-6551 + royalty extensions (like EIP-2981, or custom split contracts)
Access Control
Tokenbound accounts (TBA) + permission logic (EIP-5192)
Endpoint Enforcement
Gateway middleware signs all usage, logs back to chain (or L2)
Retraining Logic
Contributors can vote to “upgrade” model, replace checkpoint hash, and fork as new M-NFT
Revenue Streaming
Superfluid / Sablier compatible flows for real-time inference fee streaming
Governance Over Model
Optional M-NFT-level subDAO (based on token delegations or voting weights)
🔁 Lifecycle of an M-NFT
💰 Revenue Distribution Example
Suppose a model is co-trained by 3 parties:
Alice (architecture + base training): 50%
Bob (fine-tuning specialist): 30%
DAO Treasury (compute subsidy): 20%
Every time the model is called:
API usage is logged → signed by gateway
User pays in $ASTRA → smart contract verifies + routes funds
Each party receives automatic share (instant or streaming)
✅ No need for off-chain SaaS billing — protocol-native inference economy.
🧬 Model Forking & Composability
Models can be forked into new NFTs with different checkpoints (e.g., further fine-tuned versions)
Child M-NFTs can opt to stream revenue back to parent (inheritance or upstream royalty)
Models can be composed into multi-agent systems or chained into larger AI workflows (graph execution)
M-NFT becomes the "ERC20 of AI intelligence": composable, traceable, programmable.
🧩 Future Enhancements (Planned in ASTRAI Phase 3+)
zkML Proofs: On-chain attestation of inference results via zkML circuits
Agent-bound NFTs: Combine model + behavior logic for on-chain autonomous agents
Model Fusion Logic: Cross-model ensembles with shared revenue trees
On-chain Reputation Score: Based on usage, accuracy, feedback loops (feedback-to-earn)
Fully On-chain Inference (L2 Optimistic / zk Compute)
🚀 Why M-NFT Matters
AI models are no longer just weights and files — they are programmable, composable, value-generating digital primitives.
With M-NFT, ASTRAI unlocks:
Permissionless AI economies
Model contributor monetization
Shared governance over critical AI infrastructure
Crypto-native composability for AI tooling
Last updated