Download
FlashProve is an open-source research prototype developed by Anthropic for generating zero-knowledge proofs of AI model inference. It allows an AI provider to demonstrate that a response came from a specific committed model without revealing the model's weights. The system uses cryptographic commitments to bind inference results to a defined model configuration. GPU-based proving generates evidence that can be independently checked using a lightweight CPU verifier. Its implementation supports experiments with Llama 4 Maverick, including full-size inference on an NVIDIA H100 GPU. The protocol's security properties are formally verified using Lean 4. FlashProve includes example proofs, testing utilities, and reproducible research documentation. The repository is a research artifact and is no longer maintained.
Features
- Zero-knowledge cryptographic proofs verifying AI inference against committed model weights
- Confidential model weight commitments without exposing proprietary parameter information
- GPU-accelerated proof generation with independent CPU-based verification
- Experimental Llama 4 Maverick inference proving on NVIDIA H100 hardware
- Formally verified protocol soundness using Lean 4 mathematical proofs
- Reproducible research workflows with example proofs, test suites, and verification utilities
Project Samples
