Use cases · Decentralized inference

On a shared inference network, keep a record of who served what.

On a peer-to-peer inference network, a request may be served by a GPU you've never met. Afterwards, neither the requester nor the node has a record the other can check of which model served it and what was returned.

Request a demo → See what runs online

What gets recorded

  • Each exchange the node serves, sealed on that machine by the capsule-emit-mesh plugin.
  • A checkpointed local log. Only a checkpoint hash goes to the witness, and only if you set a witness URL.
  • A per-request client nonce (mesh-llm #1397, in release v0.76.2), so a record ties to one request.

What anyone can check

  • Anyone holding a record can verify it offline and check its checkpoint against the public witness.

What runs today

What's built, and what isn't yet.

  • Runs today The plugin, v0.1.2: Mesh-LLM/capsules v0.1.2 ↗.
  • Preview Weights digest and serving provenance on each response (mesh-llm #1708, #1841): merged on mesh-llm main, not yet in a release.
  • Preview Confirmation from the other side: joining the requester's half to the node's half by exchange id works; the acknowledgment leg is not built yet.
  • Preview An Evidence tab in the mesh-llm console. It is not in a released mesh-llm yet.
# after downloading and checking the release archive
mesh-llm plugins install --archive capsule-emit-mesh-0.1.0-<target>.tar.gz \
  --name capsule-emit-mesh --version 0.1.0

mesh-llm on GitHub ↗