Anthropic Releases FlashProve: Provable Inference Tool for MoE Models

Anthropic releases FlashProve, a cryptographic verification system for large language model outputs

A collaborative effort between Anthropic’s Frontier Red Team and researchers at EPFL, MIT, and StarkWare has resulted in the open-sourcing of FlashProve, a system designed to generate cryptographic proofs confirming that specific language models produced particular responses. The project includes a prover, two verification tools, formal mathematical proofs, and a protocol specification, all available on GitHub under the Apache 2.0 license.

Addressing model deployment challenges

The system addresses a critical challenge in model deployment: verifying that a response originates from a specific model rather than an alternative or tampered source. When interacting with an API, users cannot confirm which model processed their request. Providers might route queries to lower-cost models, while unauthorized access to inference servers could alter token streams without detection. FlashProve enables independent verification of model authenticity, ensuring responses align with the expected model’s output.

Core components of FlashProve

The repository contains three core components: a prover, a high-performance verifier implemented in C++, and a secondary verifier built using Lean, a proof assistant. The Lean-based verifier serves as a reference implementation, rigorously validating the protocol’s completeness, soundness, and zero-knowledge properties. While the C++ verifier is optimized for practical use, the Lean version operates more slowly but provides mathematical assurance.

Verification and manual checks

The team acknowledges potential discrepancies between the paper, formal specifications, and codebase, requiring manual verification of Lean statements. An interactive visualizer is included to aid this process. The prover itself lacks formal tooling, necessitating traditional auditing methods. The system does not include model weights, which users must obtain separately from Hugging Face under Meta’s licensing terms.

Testing and performance metrics

FlashProve was tested on Llama 4 Maverick, a 400-billion-parameter Mixture-of-Experts (MoE) model. Execution requires an 80 GB GPU with 290 GB of host memory, as the prover processes only 17 billion active parameters per token, streaming weights in layer-sized segments. Performance metrics demonstrate scalability: on an H100 GPU with 24 CPU cores, the prover handles approximately 110 tokens per second. A 1,024-token request generates a 18.78 MB proof, which verifies in 17 seconds using 24 CPU threads or 21 seconds on a single thread. The initial offline phase, where the prover commits to model weights, takes 39 minutes and supports roughly 16 million requests.

Advancements in model transparency and security

The project highlights advancements in model transparency and security, offering a framework to validate inference outputs in environments where trust in infrastructure is critical. By combining cryptographic proofs with formal verification, FlashProve provides a foundation for auditing large-scale language model deployments.

According to the research team, FlashProve represents a significant step toward ensuring model authenticity and accountability in AI systems.