A decentralized AI network for people who have to keep it running.
Most decentralized-AI projects ship a consensus paper and a whitepaper. SolanaLM ships those plus the FastAPI gateway, the Prometheus integration, the SQLAlchemy migrations, the Docker Compose file, the Kubernetes manifests, the Textual TUI, and the audit log schema. Because that's what running a real network looks like.
Our thesis
The agent economy will make more inference calls than humans ever did, at thinner margins, with agents that increasingly hold their own wallets. Centralized providers will keep getting cheaper, but they will never offer per-request, machine-to-machine economics for those workloads — and they will never let you verify that your prompt is unreadable to the operator or that the model you paid for is the one that ran. That needs a decentralized substrate settling on-chain: Solana's sub-second finality and sub-cent fees finally make per-inference settlement practical, and an on-chain audit trail is the first step toward verifiable inference.
Why Solana specifically
We need block finality faster than a human reads the first token of a response. We need fees lower than the marginal cost of an inference call. Solana hits both: ~400ms finality, ~$0.00025 per transaction. Polkadot subnets, Ethereum L2s, and Cosmos chains all forced compromises. Solana didn't.
What we shipped
v0.1.0 is small but complete. You can run the gateway and three nodes locally with three commands, and the same code paths run in production behind a Kubernetes load balancer.
Gateway
FastAPI + Uvicorn + Pydantic v2. Auth, rate limits, request routing, audit logs.
SDK
Python client with async + sync APIs and an OpenAI-compatible shim.
Inference runtime
PyTorch/Transformers, llama.cpp via llama-cpp-python, proxied OpenAI/Anthropic/Cohere/Ollama.
Federated training
FedAvg, FedProx, FedAdam, SCAFFOLD, with secure aggregation and differential privacy.
Settlement
solana-py + pynacl + base58 for signed transactions. Per-request micropayments.
Operations
Prometheus, Redis cache, asyncpg + SQLAlchemy 2, Alembic, Textual TUI, WebSocket admin.
Operating principles
Boring is a feature
Postgres, Redis, FastAPI, Prometheus. Standard stack. The novelty is in the protocol, not the bill of materials.
Compatibility wins
OpenAI's schema is a public good. We implement it instead of inventing a new one.
Privacy is default-on
Onion routing and differential privacy ship in the box, not as plugins someone might enable later.
Open source, MIT
Fork it, audit it, run a competing registry. We'll take the better protocol, not the higher fee.
Team
cryptuon is a small group of engineers shipping production crypto-native infrastructure. We previously worked on payments, distributed systems, and ML platforms. We use the tools we ship.