Anthropic open-sources FlashProve for verifiable model inference
Anthropic open-sourced FlashProve, a zkSNARK-based system that lets model providers prove a response came from a specified model without revealing its weights. In testing on Meta's 400B-parameter Llama 4 Maverick, one NVIDIA H100 generated proofs for about 110 tokens per second, with roughly 19 MB per 1,024-token request; a 24-thread CPU took about 17 seconds to verify one proof. Anthropic also released the proof generator, C++ and Lean verifiers, documentation, and formal proofs under Apache 2.0, but the current prototype supports 1,024-token contexts, integer models, and greedy decoding, while proof generation requires roughly 70–95 times ordinary vLLM GPU time. FlashProve verifies model identity and inference integrity, not the factual correctness of an answer. Why it matters: FlashProve offers a cryptographic way to audit model identity and output integrity without disclosing model weights, although its current proof cost and limited inference support constrain broad use. What do you think?
Open the headline: source, AI briefing and more →Comments
…