Anthropic’s FlashProve Generates Llama 4 Proofs at About 110 Tokens Per Second

Anthropic released FlashProve on Nov. 2, a research prototype that uses cryptographic proofs to check model execution, with a Llama 4 Maverick benchmark reaching 110 tokens per second.

The benchmark generated proofs at about 110 tokens per second on one H100 GPU. Each proof was about 19 megabytes for a request containing 1,024 tokens. Llama 4 Maverick has 400 billion parameters and uses a mixture-of-experts design. The test used 4-bit quantized weights and integer arithmetic during prefill.

The prototype uses zkSNARKs, a form of zero-knowledge proof that lets a verifier check a computation while keeping model weights private. The proofs are meant to help confirm which model handled an inference request, including when a customer uses a third-party routing service or an evaluator checks a model.

It checks execution under a defined setup; it does not assess whether an answer is accurate, safe or unbiased.

FlashProve’s proving time was about 70 to 95 times the GPU time of running the model with vLLM.

The release includes a research post, technical report, protocol description, executable code with a prover and two verifiers, and Lean proofs for top-level mathematical statements. It also includes an interactive visualizer for reviewing the theorems. Anthropic released the code, Lean proofs and documentation under the Apache 2.0 license and described FlashProve as a progress update. The proof addresses model execution, while answer quality remains outside its scope.