Anthropic releases FlashProve, a zero-knowledge proof system for verifying LLM inference

·2026.11.10 02:47

Key point

The Apache 2.0 research prototype verifies that a specific model generated a response using zero-knowledge proofs, with a demo running Llama 4 Maverick at 110 prefill tokens/sec.

Details

Anthropic released FlashProve, an Apache 2.0 research prototype that uses zero-knowledge proofs to verify LLM inference without exposing model weights.
The system allows anyone to verify a proof on a regular CPU in seconds, confirming the provider committed to a specific set of weights for a given reply.
A demo runs Llama 4 Maverick at full size on a single H100 GPU, achieving approximately 110 prefill tokens/sec in batches of 24.
Protocol soundness is machine-checked in Lean 4, but the repo is currently marked as "not maintained."
Current limitations include proof sizes of approximately 18.8 MB, verification of only one output token per proof, and a context cap of 1,024 tokens.

security open-source anthropic llm-inference zero-knowledge-proofs flashprove

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.