Overview
- Early November Anthropic published FlashProve as an open-source research prototype that produces cryptographic, zero-knowledge proofs tying a specific output to a committed model fingerprint.
- FlashProve works in two steps: labs first publish a small model commitment that fingerprints weights, then the system generates a per-response zkSNARK that anyone can check without access to the model itself.
- Benchmarks on a 400B mixture-of-experts model showed proof generation at about 110 tokens per second on one NVIDIA H100, proofs of roughly 19 MB per 1,024 tokens, and CPU verification in about 17–21 seconds.
- The prototype is not production-ready because proving costs are high — Anthropic reports roughly 70–95× the GPU time of standard inference — and it currently supports limited modes (1,024-token contexts, integer/greedy decoding and partial floating-point coverage).
- Anthropic released code, executables, a protocol spec and a Lean formalization under Apache 2.0 to invite academic and community work to cut proving overhead, broaden supported inference modes, and enable audits and regulator checks that confirm model provenance.