Weights are content addressed
GGUF files are chunked by transformer block, hashed with BLAKE3 into CIDv1, and moved over a 64 MiB sliding window on libp2p request-response. No model host to take down.
02Inference network
A peer-to-peer inference network that shards a model across consumer hardware. Transformer blocks are distributed by pipeline parallelism, hidden state tensors move over a low-latency mesh, and each contribution is settled against an on-chain attestation rather than an invoice.
Supplies compute. Contributors earn against Proof of Contribution, verified off chain and anchored on chain.
GGUF files are chunked by transformer block, hashed with BLAKE3 into CIDv1, and moved over a 64 MiB sliding window on libp2p request-response. No model host to take down.
Federated training uploads weight gradients only. Contributor data does not leave the contributor device.
Rewards are escrow funded. There is no mint and no inflation. The chain enforces eligibility, maturity, disputes, and refunds.
Values are taken from the protocol repository and from live chain parameters.
OmniNode specification
| Sharding | Pipeline parallelism across transformer blocks |
|---|---|
| Weight distribution | BLAKE3 to CIDv1, 64 MiB sliding window over libp2p |
| Zero-knowledge proofs | Halo2 circuits |
| Signing | Ed25519 over BLAKE3 digests |
| Attestation domain | omninode.inference_attestation.v1 |
| Activation height | 6,000,000 |
| Claim maturity | Inclusion plus finality depth plus dispute window |
| Dispute resolution | Validator quorum, basis-point threshold, no resolver key |
| Slashing in v1 | None. Levers are reward denial, claim withholding, escrow refund. |
| Privacy | Federated learning, gradient upload only |
| Read RPC | sum_getInferenceAttestation, sum_listInferenceAttestations |
| Toolchain | Rust edition 2024, toolchain 1.85 |
| License | MIT or Apache-2.0 |