The prosperity of the open-source AI ecosystem brings convenience, but it has also given rise to an endless stream of "wrapper models"—some developers or organizations lightly repackage open-source models, or even swap weights and tamper with outputs, then claim them as entirely new, self-developed models. To tackle this industry chaos at its root, leading AI company Anthropic recently officially launched and open-sourced a new provable inference system for large models—FlashProve. Using cryptography at its core, the system can precisely verify a model's true identity, setting a new benchmark for transparency and compliance in the large-model ecosystem.
Technical deep dive: a three-step verification loop based on zkSNARK
As a system focused on "verification," FlashProve's core technical approach abandons the traditional cumbersome process of obtaining massive model weights or rerunning the entire model. Instead, it innovatively introduces the zero-knowledge proof technology zkSNARK. The entire verification process is streamlined into an efficient three-step loop:
Offline generation of a unique weight commitment ID: Before a model is deployed, the system generates a unique cryptographic commitment identifier (Commitment ID) as the model's tamper-proof "digital ID card."
Online generation of a cryptographic inference proof: When a large model processes user requests in the cloud and produces results, the system simultaneously generates the corresponding cryptographic inference proof.
Fast 17-second verification on an ordinary CPU: The verifier does not need to retrieve the model's expensive weights or rerun the model. On an ordinary CPU terminal, it takes only 17 seconds to fully verify the cryptographic proof.
