88Z AI, 4 November — Anthropic has open-sourced FlashProve, leveraging zero-knowledge proofs to allow service providers to verify that responses originate from a specified model without exposing weights. Tested on Llama 4 Maverick, the system enables trusted verification despite higher costs.
At a glance
- FlashProve uses zero-knowledge proofs to verify model origin without leaking weights
- Proof costs are 70-95 times higher than standard inference in Llama 4 Maverick tests
The update
Anthropic has officially open-sourced FlashProve, a verifiable reasoning system. At its core, the system applies zero-knowledge proof technology, allowing service providers to prove to users that generated responses indeed come from a specified model architecture while strictly protecting model weight privacy, thereby establishing a trust mechanism. Currently, FlashProve has been tested on Llama 4 Maverick, a mixture-of-experts model with 400 billion parameters. Data shows that on H100 hardware, its inference speed is approximately 110 tokens per second. For every 1024 tokens generated, the resulting proof size is about 19MB, with CPU verification taking around 17 seconds. Notably, the current cost of generating proofs is approximately 70 to 95 times that of standard inference, indicating that the technology still faces high computational overhead challenges in practical deployment.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 机器之心Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.