Anthropic Open-Sources FlashProve for Zero-Knowledge Model Provenance

By ▲ Marktechpost AI (@Marktechpost)Original title: ▲ Anthropic just open-sourced FlashProve, a way for an AI provider to prove which model actually ans...
85AI Score

Anthropic released FlashProve, an open-source tool using zkSNARKs to prove which model generated a response without revealing weights. It achieves ~110 tokens/sec on Llama 4 Maverick, thousands of times faster than VerInf, though at the cost of 70-95x inference compute and limited context length.

Full Report

▲ Marktechpost AI · 2026-11-03 22

▲ Anthropic just open-sourced FlashProve, a way for an AI provider to prove which model actually answered your request, without revealing the model weights.

The key idea is a zkSNARK proof attached to each reply. The provider first publishes a short commitment that binds 1 exact set of weights. Each proof then shows those committed weights produced these exact tokens, and anyone can check it on a CPU.

On Llama 4 Maverick (400B MoE, ~17B active), a single H100 proves about 110 tokens/second. Proofs are 18.78 MB and verify in 17 seconds on 24 threads. The prior open-source proof of the same model, VerInf, ran at about 0.02 tokens/second. It is a trade-off though: proving costs about 70 to 95x the GPU time of plain vLLM inference, and context is capped at 1,024 tokens.

Full analysis:

推文配图