Anthropic’s FlashProve produced verifiable execution proofs for Llama 4 Maverick at about 110 tokens per second, offering a way to check model use while consuming far more GPU time.
On one H100 GPU, the prototype generated verification material at roughly 110 tokens per second. A proof for a 1,024-token request measured around 19 MB. Creating proofs took 70 to 95 times as much GPU time as the vLLM run.
For prefill, the benchmark used four-bit weights and integer calculations. Maverick pairs 400 billion parameters with a mixture-of-experts architecture.
FlashProve relies on zkSNARKs, which let a checker validate a computation without receiving the model weights. That can help a customer or evaluator establish which model ran a request, including when a third-party service routes it. The check establishes conformity with a specified configuration; it does not assess whether the resulting answer is accurate, safe or unbiased.
Anthropic’s release contains prover code, two verifier programs, Lean proofs of its main mathematical claims, documentation and an interactive theorem viewer. The code, proofs and documentation are available under the Apache 2.0 license.
