Overview
- Anthropic published FlashProve as an open-source research prototype under Apache 2.0 that the company released on Tuesday, November 3, 2026.
- The system requires an offline cryptographic commitment or fingerprint of model weights and then produces a per-inference hash-based zkSNARK that anyone can verify without seeing the weights.
- Anthropic demonstrated the approach on Meta’s 400B-parameter Llama 4 Maverick and reported roughly 110 tokens per second on one NVIDIA H100, proofs of about 19 MB per 1,024 tokens, and CPU verification times around 17 to 21 seconds.
- FlashProve currently carries large practical costs and limits: proving uses about 70–95 times the GPU time of standard vLLM inference, supports only up to 1,024-token contexts, and proves an integer model under greedy decoding rather than full floating-point or sampled outputs.
- Anthropic and contributors published code, a protocol spec, and machine-checked formal proofs to invite researchers to cut proof size and compute overhead so the technique can be used for audits, anti-tampering checks, and regulatory verification.