[Anthropic Open-Sources FlashProve, Enables Zero-Knowledge Proofs for Large Model Inference with a Single H100 GPU] According to TechFlow, on November 3rd, the Anthropic Frontier Red Team released the FlashProve system, exploring the use of zero-knowledge concise non-interactive knowledge proofs (zkSNARK) to verify the reasoning process of large language models. This allows users to verify whether an answer was indeed generated by a specified model without obtaining model weights. Tests show that FlashProve already supports the Llama 4 Maverick hybrid expert model with 400 billion parameters. Under 4-bit quantization, a single NVIDIA H100 GPU can achieve a proof generation speed of approximately 110 tokens/s, generating approximately 19 MB of proof files per 1024-token request. Users can complete verification in approximately 17 seconds using the CPU, without requiring a GPU. Compared to running the model directly using vLLM, generating proofs requires approximately 70 to 95 times more GPU time. Anthropic states that this technology has the potential to be applied to model authentication, tamper-proofing of the inference process, AI regulatory compliance, and model service auditing, preventing service providers from replacing models without authorization or attackers from tampering with model output. Currently, FlashProve is still a research prototype and has limitations such as context length restrictions and differences between floating-point inference and integer model proofs. Anthropic has also released the relevant code, license specifications, and Lean formal verification materials, open-sourced under the Apache 2.0 license.