Anthropic Announces FlashProve: Zero-Knowledge Proofs That an AI Reply Came From Committed Model Weights

Anthropic's Frontier Red Team has released FlashProve, an open-source research prototype for provable inference. FlashProve is a zkSNARK system that proves a reply came from a committed model without revealing anything about the model weights. On Llama 4 Maverick, a 400B mixture-of-experts model, it sustains proving throughput of roughly 110 tokens per second on a single H100. For developers, it addresses a persistent gap: today you cannot check which model actually answered your API call.

TL;DR

• Size: Proves a 400B MoE model (~17B active parameters per token); context is capped at 1,024 tokens per request.
• Runs on: The prover needs one H100 80 GB GPU, 24 CPU cores, and ~290 GB of host RAM. The verifier runs on a CPU alone—no GPU required.
• Performance: Orders of magnitude faster than prior open-source zkML systems on the same 400B model.
• Best case: 110 tokens/second proving on a single H100, with 18.78 MB proofs checked in 17 seconds.
• Worst case: Proving costs roughly 70–95x the GPU time of plain vLLM inference.

What Problem Does Provable Inference Solve?

A user querying a model cannot verify which model production actually served. To help close that gap, FlashProve lets a provider publish a commitment to its model weights and then generate a cryptographic proof that a given response was produced by those exact weights—all without exposing the weights themselves. The 2026 release arrives as enterprises, regulators, and API consumers increasingly demand verifiable model provenance rather than trust in provider attestations alone.

via MarkTechPost