Anthropic Open-Sources FlashProve: Zero-Knowledge Proofs for Large Model Inference on a Single H100

On November 3, according to Anthropic's technical blog, the Anthropic Frontier Red Team has launched the FlashProve system, exploring the use of zero-knowledge succinct non-interactive knowledge arguments (zkSNARK) to verify large language model reasoning processes, enabling users to confirm that responses were genuinely generated by a designated model without obtaining its weights.

Tests show that FlashProve now supports the 400-billion-parameter Llama 4 Maverick Mixture of Experts model. Under 4-bit quantization, a single NVIDIA H100 GPU achieves a proof generation speed of approximately 110 tokens/s, producing about 19 MB of proof files per 1024-token request. Users can complete verification via CPU in roughly 17 seconds without needing a GPU. Compared to running the model directly using vLLM, generating proofs requires approximately 70 to 95 times more GPU time.

Anthropic stated that this technology holds potential applications in model authentication, tamper-proof reasoning processes, AI regulatory compliance, and model service auditing, preventing providers from substituting models without notification or attackers from tampering with model outputs.

Currently, FlashProve remains a research prototype and faces limitations such as context length restrictions and discrepancies between proofs for floating-point and integer models. Anthropic has simultaneously released the related code, protocol specifications, and Lean formal verification materials open-sourced under the Apache 2.0 license.