The TrueOpen team has launched a public testnet for an open network that enables verifiable and private AI inference using independent GPUs. The system ensures that the agreed-upon model was actually executed while keeping inputs and outputs encrypted.

  • Logprob Majority-Consensus Verification (LMCV) allows independent verifiers to recompute output-token logprobs, ranks, and top-k sets via teacher-forced prefill.
  • Sampled Layerwise Proofs (SLP) commit to all computation chunks but only prove mandatory ones plus a random sample to reduce costs.
  • In tests with Qwen3-8B and 32B models, verification cost was approximately 1 to 8% of generation GPU time.
  • Node operation and model registration are now rolling out on the testnet, with settlement handled via USDC.

The team is inviting community scrutiny of the verification approach and welcomes feedback from those running GPUs or maintaining open models.