A decentralized network for private, censorship-resistant AI inference — where no single entity controls the model, sees your data, or decides what you can ask.
Centralized AI providers filter, log, and refuse requests at will. UNFED AI distributes inference across independent operators — no single entity can censor or shut down the network.
Your prompts never exist in plaintext on any single machine. Multi-Party Computation at the embedding layer means the most sensitive part of inference — turning your words into numbers — is cryptographically protected.
Models are sharded across multiple nodes run by different operators. If one node goes down, the network can route around it. No company can pull the plug.
The fundamental difference in who controls your data.
The embedding layer (shard 0) runs as an MPC pair. Your token IDs are secret-shared before any computation happens. Subsequent shards only see intermediate activations, not your original text.
Nodes must stake tokens on-chain to participate. A verifier network randomly spot-checks computations. Cheating results in automatic slashing — dishonesty costs real money.
Registries (clusters) compete like Monero mining pools. Each operator sets their own pricing, staking rules, and model offerings. Nodes and clients choose freely based on economics and reputation.