Anthropic Releases FlashProve for Cryptographic Verification of LLM Inference
Frontier AI models · Enterprise AI
Anthropic introduced FlashProve, an open-source prototype using zkSNARK cryptographic proofs to verify that an AI model output was generated by specific model weights without exposing proprietary parameters. For a 400B-parameter MoE model, a single H100 GPU generates proofs at approximately 110 tokens per second, with proof sizes around 19 MB per 1,024-token request.
Anthropic, in collaboration with researchers from MIT, EPFL, StarkWare, and Succinct, has published FlashProve, an open-source prototype demonstrating provable inference for large language models. The software-only approach relies on zkSNARKs (zero-knowledge succinct arguments of knowledge) to let providers prove a specific committed model generated a response without disclosing model weights. Benchmarked on a 400B-parameter mixture-of-experts model (Llama 4 Maverick, 4-bit quantization), a single H100 GPU generated proofs at roughly 110 tokens per second, taking about 70 to 95 times the GPU time of running model inference on vLLM. Anthropic released executable prover and verifier code, a formal Lean specification, and technical documentation under the Apache 2.0 license.
Key takeaways
- FlashProve provides a zero-knowledge cryptographic mechanism that allows verification of inference authenticity while keeping proprietary weights hidden.
- Benchmarking on a 400B MoE model demonstrated a generation speed of roughly 110 tokens/second on an H100 GPU, generating ~19 MB proofs per 1,024-token request.
- Anthropic released an open-source package under the Apache 2.0 license, including C++ and Lean verifiers alongside formal mathematical proofs.
- The approach addresses security and governance use cases, including model tampering detection, regulatory pacing compliance, and provenance verification.
Source：Anthropic Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.