LLUMA.BODEGGA.NET
LLUMA-SPEC-001 · REV 2 · 2026-07
Specification · anonymous inference network

Inference you can use. Identity no one can follow.

LLUMA is a torrent-style network for running language-model inference across volunteer hosts. Prompts and identity enter the network on opposite sides and meet only inside an untraceable core: the host that serves your content never learns who asked, and the host that knows who asked never sees the content. No accounts. No logs worth seizing. Zero external requests on this page — view source.

$read the protocol $run a host
THE VOID NO PATH CROSSES CONTENT HOST sees: ciphertext IDENTITY HOST sees: no content HOST 03 HOST 04 HOST 05 HOST 06
FIG. 01 Hosts communicating through the anonymous void Links dissolve. No host can trace another through the core.

§ 01Architecture

LLUMA-SPEC-001 · REV 2

Three roles, strictly separated. Anonymity is not a feature bolted on top — it is the consequence of no single party ever holding both halves of a request.

RoleFunctionSeesNever sees
RELAY Routes encrypted frames between peers. Interchangeable, unprivileged. sees: ciphertext content · identity
BROKER Matches inference demand to host capacity. Holds session tokens, not sessions. sees: capacity, tokens prompt content · client address
HOST Executes inference on volunteered hardware. Paid in network credit. sees: plaintext prompt who asked · origin, account, address
TABLE 01 — Role separation. Compromise of any one role yields half a secret.

§ 02Operation

LLUMA-SPEC-001 · REV 2

One request, end to end. Order matters; the sequence below is the protocol.

  1. Client splits the request. Identity material goes to the identity path; the prompt is encrypted to the content path. The two halves share only a one-time pairing token.
  2. Broker matches capacity. A host advertising free compute is assigned by token. The broker never receives the prompt and never learns the client's address.
  3. Relays carry ciphertext. Frames hop across volunteer relays torrent-style. Any relay sees traffic; no relay sees meaning.
  4. Host runs inference. The prompt decrypts only on the assigned host. Output returns along a fresh relay path — never the inbound one.
  5. The pairing token burns. After delivery the token is destroyed. There is no durable record linking identity to content — nothing to subpoena, nothing to leak.

§ 03Trust model

LLUMA-SPEC-001 · REV 2

Two tiers, honestly labeled. The Open tier is not zero-knowledge — the executing host sees your plaintext prompt without knowing who sent it. If host-blindness is a requirement, use the Confidential tier.

TierExecutionHost seesGuarantee
OPEN Any volunteer host plaintext prompt, anonymous origin Unlinkability — content cannot be traced to identity
CONFIDENTIAL tier: tee — attested enclave hosts sees: ciphertext Unlinkability + host-blindness, verified by remote attestation
TABLE 02 — What each tier actually promises. No tier is marketed beyond its proof.

§ 04Contribution

LLUMA-SPEC-001 · REV 2

The network is its volunteers. Contribute what you have; earn network credit against future inference.

ContributeRequirementEarns
RELAYStable uplink, any hardware★★★
BROKERHigh-availability node, staked credit★★★
HOST — OPENGPU with ≥16 GB VRAM★★★
HOST — TEEAttestation-capable enclave hardware★★★
 TABLE 03 — Contribution tiers. Credit is spendable on inference; it is not a token, coin, or security.

$lluma host --join