Anthropic releases FlashProve for cryptographic AI inference proofs

FlashProve uses zkSNARKs in a software-only design. During an offline phase, the provider publishes a short commitment identifying exact weights while keeping them secret. For completed prompt-response pairs, the system produces a proof that output tokens match the committed model's output. Anyone with the commitment, prompt, and reply can verify without accessing weights. The implementation leverages Meta's Llama 4 Maverick's mixture-of-experts structure, with prover work scaling to roughly 17 billion active parameters per token rather than all 400 billion.

Anthropic evaluated on a 4-bit conversion of Maverick using one H100 and 24 CPU cores. A batch of 24 requests took about 223 seconds to prove, producing 18.78 MB proofs each. The offline commitment step took about 39 minutes. Verification runs on CPU without GPU: 17 seconds with 24 threads, 21 seconds with one thread. Proving required about 72 times more per-GPU time than FP8 inference on eight H100s running vLLM. The repository includes a prover, C++ and Lean verifiers, and formal proofs under Apache 2.0, labeled version 0.16.0 as an unmaintained research artifact. Known limits include 1,024-token contexts, no independent security audit, and coverage only for exact integer weights under greedy decoding.