Integrating machine learning (ML) models into decentralized applications (dApps) is becoming an increasingly urgent task as interest in Web3, DePIN, and AI infrastructure grows. One of the main challenges in this area is how to efficiently and transparently store and use weighted ML models without losing the principle of decentralization. This is where IPFS comes in – a distributed file system that allows data to be stored and addressed outside the blockchain, but with full compatibility with its principles.
Storing models on IPFS solves several problems at once. Firstly, machine learning models often have an impressive size – from hundreds of megabytes to gigabytes. It is technically impossible and uneconomical to place them directly in a smart contract. IPFS allows you to load a model as a file, fix its hash (CID) and access it using this identifier. The smart contract, in turn, can reference the CID, retaining control over the version and access point of the model without having to store its contents in the blockchain.
Second, this approach enhances verifiability. Every time a dApp user interacts with the model, the system can ensure that the exact version pointed to by the contract is loaded. If someone tries to tamper with the model, the hash will change and the contract will reject it. This is especially important for use cases where model trustworthiness is important: voting, decentralized oracles, collateral evaluation, etc.
The integration is as follows: a developer trains an ML model in a traditional environment (e.g., Python + PyTorch), then serializes it and publishes it on IPFS (via a web interface or pinning service such as Pinata). The CID of this model is stored in a smart contract, which can be part of the dApp. The dApp frontend, having received the CID, loads the model in the browser or on the client, performs the inference locally and sends the result back to the contract. This reduces the load on the network and preserves data privacy.
In practice, this solution is already used in projects related to decentralized risk assessment, NFT content generation, voting systems with data preprocessing, AI games and DAOs. The key advantage is that even very complex logic can be partially moved outside the smart contract, while remaining verifiable and tied to the source code via the IPFS CID.
It is important to note that while IPFS does not provide computation (unlike zkVM or specialized networks like Gensyn, Modulus or Bittensor), it does an excellent job of solving the problem of long-term, decentralized and transparent model storage. Further steps in this direction include linking IPFS to computational markets, e.g., using off-chain inference platforms or zk proofs to prove the correctness of model execution.
Integrating ML models into dApps via IPFS is an intermediate but important solution that bridges the world of Web2 models and Web3 infrastructure. It not only reduces the load on the blockchain, but also opens up new scenarios for using AI in decentralized systems, making them accessible, repeatable, and independent of centralized APIs.
