Anthropic: Progress on provable inference for large language models

Provable inference aims to let users verify that an AI response came from a particular model without accessing its private weights or rerunning the computation. Anthropic presents FlashProve, an open-source prototype using zero-knowledge succinct arguments (zkSNARKs): a provider commits to a model and supplies a compact proof with each response. The proof can be checked cheaply and reveals no information about the weights, though it does not conceal the model’s architecture. The approach addresses a basic trust problem: output quality alone cannot show whether a service used the promised model, a cheaper substitute, or a model whose responses were tampered with.

FlashProve demonstrates proof generation for Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model, using 4-bit quantized weights. On one H100 GPU it produces proofs at about 110 tokens per second, with a roughly 19 MB proof for a 1,024-token request; proving takes about 70-95 times the GPU time of ordinary inference with vLLM. The work argues that such proofs could help customers detect substitutions, help organizations audit model use, and support regulators checking claims about compute devoted to inference - although proofs alone cannot show that hardware is not also training. The release includes code, protocol documentation, and Lean formalizations and proofs of key security statements, published under Apache 2.0. The project is presented as an intermediate, auditable contribution intended for others to improve.