Anthropic Open-Sources FlashProve to Enable Zero-Knowledge Proofs of Large Model Inference on a Single H100

DeepTechFlow reported that, according to Anthropic's technical blog on November 3, the Anthropic Frontier Red Team launched the FlashProve system, exploring the use of zero-knowledge succinct non-interactive arguments of knowledge (zkSNARKs) to verify large language model inference processes, enabling users to verify whether responses were indeed generated by a specified model without accessing model weights.

Testing shows that FlashProve already supports the 400-billion-parameter Llama 4 Maverick mixture-of-experts model. Under 4-bit quantization, a single NVIDIA H100 GPU can achieve proof generation at approximately 110 tokens/s, producing about 19 MB of proof files for each 1024-token request. Users can complete verification on a CPU in about 17 seconds without a GPU. Compared with running the model directly using vLLM, the GPU time required to generate proofs is approximately 70 to 95 times greater.

Anthropic said the technology could be applied to model authentication, tamper-proofing inference processes, AI regulatory compliance, and model service audits, preventing service providers from replacing models without notice or attackers from tampering with model outputs.

FlashProve remains a research prototype, with limitations such as context-length constraints and differences between floating-point inference and integer model proofs. Anthropic has also released the related code, protocol specifications, and Lean formal verification materials under the Apache 2.0 license.

Source: DeepTechFlow