🤖 Anthropic releases FlashProve — verifiable model inference
The Frontier Red Team has released a FlashProve prototype (v0.16.0, Apache 2.0) based on zkSNARKs: the provider publishes a commitment hash of the weights without revealing the weights themselves, and attaches a proof to the response that the tokens were generated by this specific model. Verification takes approximately 17 seconds on a 24-core CPU; for Llama 4 Maverick (400B MoE), the hash is approximately 6.5 MB and covers about 16 million requests.
🌍 API providers will be able to prove to clients that a response was generated by the stated model, not a cheaper substitute, which addresses trust issues in LLM traffic routing and provides regulators with a tool to verify hardware.
👤 This is an open prototype, not a service: the verifier compiles on Linux (g++ 13+, <2 GB RAM) in a couple of minutes. Note: the verifier receives the request and response themselves, so privacy must be evaluated separately.
Source 1: https://www.anthropic.com/research/provable-inference
Source 2: https://github.com/anthropics/FlashProve