anthropics/flashprove is a research prototype for zero-knowledge proofs that show an AI model's responses came from a committed model. A zero-knowledge proof demonstrates that a condition holds without revealing the content itself. This repository covers a design that verifies the origin of a response without publishing the weights. The GitHub repository description states that it is not maintained.
The model provider publishes a commitment identifier tied to the weight bundle, and the GPU generates a proof that backs the origin of each response. It is designed so that anyone else can verify a proof on an ordinary CPU within seconds. The README covers the soundness of the proof protocol with Lean 4 theorems, and documents how to check the stored example proof and how to run the toy proof. It notes that running the full-size model requires a separate high-performance GPU and large memory. The repository also includes a case of running Llama 4 Maverick on a single H100. The project is written in Lean and follows the Apache-2.0 license.
However, this repository is unmaintained research material, and the documented full-size model run also requires separate hardware. So it is closer to a research record for looking at how to hide weights and prove the origin of responses than a practical tool with current support.