I’m curious to learn if anyone has attempted to use Solana to store and run a Neural Network on-chain?

Essentially having each weight stored as an account, and each layer as an account of weight-accounts and activation-accounts. then have the on-chain program process data through a series of layer-accounts.

Probably not cost effective to train a network on-chain, but once weight values have been trained (through a system such as PyTorch) could the network be de-constructed and run on-chain?

1 Answer 1


Yes. You can use Seahorse Lang to use Python code to create Solana program, and potentially even run a neural network.

In fact, somebody did this in 2022, check this out: https://sea-nn.vercel.app/

It a convolutional neutral network trained off-chain on MNIST data and the trained model is uploaded on-chain. You draw a digit, which is sent as a parameter, and then the on-chain model returns prediction output in program logs.

  • Fantastic! Thank you.
    – P-Rod
    Commented May 26 at 4:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.