Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights

| Source: MarkTechPost

Tags: Anthropic, FlashProve, zkSNARK, provable inference, Llama 4 Maverick, H100

Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that proves a model reply came from committed weights without revealing them, running Llama 4 Maverick at 110 tokens per second on one H100.

Details

According to MarkTechPost, Anthropic has released FlashProve, an open-source prototype under the Apache 2.0 license that uses zkSNARKs to prove a reply was produced by a specific committed model without exposing the weights. The report says it proves Llama 4 Maverick inference at 110 tokens per second on a single H100, with proofs of about 19 MB. For API buyers and auditors, provable inference addresses a real gap: today customers mostly have to trust that a provider served the model it claims. Proof size and throughput figures suggest the approach is moving toward practical use, though it is described as a prototype.