Anthropic Releases FlashProve for Zero-Knowledge Proofs of AI Inference
Anthropic released FlashProve, a research prototype that lets an AI provider prove which model answered a request without revealing the model's weights. The system runs Llama 4 Maverick at full size, with a single H100 GPU generating proofs at around 110 tokens per second and a proof size of about 19 MB per 1,024-token request. Anthropic open-sourced the code, Lean proofs, and documentation under the Apache 2.0 license.
Event version 1.
Key insights
Commit, prove, verify
A provider publishes a commitment ID binding it to one set of weights, which stay secret. For each reply, the provider's GPU produces a proof that the reply is the committed model's output, and anyone can check the proof on an ordinary CPU in seconds with cp_verify without the weights.
Performance on a large model
For Llama 4 Maverick, a 400B-parameter mixture of experts, a single H100 GPU generates proofs at around 110 tokens/second with a proof size of about 19 MB per 1,024-token request, using 4-bit quantized weights and integer arithmetic in prefill. This is roughly 70-95x the GPU time of running the model itself on vLLM.
Formal verification in Lean
Anthropic provided top-level auditable statements in Lean for the claimed security guarantees, along with Lean proofs for the top-level mathematical statements and an interactive visualizer of the Lean theorems to aid auditing.
Five released artifacts
Anthropic released a blog post outlining the system and its applications, a technical report, a protocol specification document, executable code including a prover and two verifiers, and Lean proofs. The code, Lean proofs, and documentation are released under the Apache 2.0 license.
Stated applications and limitations
The proofs can make it harder for attackers to hijack a chat and enable safety evaluators to verify that the model they are testing is the same one being served to customers. The commitment and inference proofs hide the model's weights, but hiding the model's architecture is not a goal of the system.