Anthropic Releases FlashProve: Provable Inference Prototype for MoE Models

Anthropic Releases Open-Source Tool for Verifying Model Outputs

Introduction to FlashProve

Anthropic’s Frontier Red Team, in collaboration with researchers from EPFL, MIT, and StarkWare, has developed FlashProve, an open-source framework designed to generate cryptographic proofs for language model outputs. Hosted on GitHub under the Apache 2.0 license, the project includes a prover, two verification systems, formal proofs, and a protocol specification. FlashProve addresses critical challenges in ensuring the authenticity of model responses in distributed inference environments.

Key Components of FlashProve

Core Components

The repository contains three core components: a prover, a high-performance C++ verifier, and a Lean-based verifier used for formal validation. The Lean verifier, though slower, serves as a reference implementation to confirm the protocol’s correctness through mathematical proofs. Both verifiers were tested against identical scenarios, with consistent results. The C++ version is optimized for practical use cases.

Technical Validation

FlashProve was validated using Llama 4 Maverick, a 400-billion-parameter Mixture-of-Experts (MoE) model. Users must obtain the model weights separately from Hugging Face under Meta’s licensing terms. The system requires an 80 GB GPU with 290 GB of host memory, leveraging a streaming approach to handle the 17 billion active parameters per token.

Performance Metrics

On a single H100 GPU with 24 CPU cores, the prover processes approximately 110 tokens per second. A 1,024-token request generates a 18.78 MB proof, which takes 17 seconds to verify across 24 CPU threads or 21 seconds on a single thread. The initial offline phase, where the prover commits to model weights, takes 39 minutes per model and supports up to 16 million subsequent requests.

Project Complexity and Challenges

The project highlights the complexity of aligning theoretical proofs, formal specifications, and operational code. While the Lean-based protocol and paper provide foundational guarantees, the team acknowledges potential discrepancies between these elements. Manual review of Lean statements is required, supported by an interactive visualizer included in the repository. The prover itself lacks formal verification tools, relying on traditional auditing methods.

“This initiative underscores ongoing efforts to enhance transparency and trust in large-scale language model deployments, particularly in scenarios where model integrity is critical.”

Conclusion

FlashProve represents a significant step forward in verifying the authenticity of language model outputs. By combining cryptographic proofs with formal validation, the tool addresses critical challenges in distributed inference environments. Its open-source nature and collaborative development further emphasize the importance of transparency and trust in AI systems.