On-Chain Models

2025. Looking at Solana thinking "what if we could run actual neural networks on this thing." Everyone thought it was nuts. Fair enough — Solana is not exactly the kind of infra where you wake up one morning and go "hey let me run AI models on a distributed immutable ledger."

But the idea wouldn't let go. First prototype shipped using SolanaPython — running a VM on a VM, layers on layers. The cost per inference and CU were insane. Cool experiment, but absolutely not what was needed to build something real at scale.

Back to the drawing board. The result: a complete suite of tools to compile Python models directly into sBPF bytecode. At one point we were thinking about bypassing the sBPF compiler entirely, but validators need their say, so we kept things clean.

Micro-models running inference for the same cost as sending an SPL transfer. Deployed for $1.

Picture this: a network of micro-models all interconnected, fully on-chain, fully verifiable on Solana. Think about what agents can build on top of that. What programs can do with it. No off-chain dependency, no trust assumptions, no API keys.

The pipeline: Python model → Graph IR → AOT compilation → sBPF bytecode → deployed as a Solana program. Every inference call is a transaction. Every output is verifiable. Every model is permissionless.

We're building a toolkit so anyone can train their own micro-model on-chain and get paid per inference for their work.

Next phase: bringing Nanochat-scale architectures fully on-chain. We've already hit models with nearly 1 million parameters, and with the new wave of AI compression techniques it's very doable to go further.

The first 100 on-chain AI model access keys — registered as 8004 assets on Solana. Each key is a unique NFT that represents early access to the on-chain model ecosystem.

These are the first 100 assets ever registered in the 8004 standard on Solana. OG-numbered #001 to #100. They don't just look cool — they're functional access keys for a new kind of infrastructure.