DeepProve is a zero-knowledge proof system for neural network inference that generates cryptographic proofs of model forward passes using sumchecks and logup GKR.
The tool addresses the challenge of proving that a neural network—particularly large language models—executed correctly without revealing the model weights or intermediate computations. It achieves sublinear proving time in model size by leveraging interactive oracle proofs rather than circuit-based approaches, enabling orders of magnitude faster proof generation than prior work. The system supports end-to-end LLM proving across transformer layers, from token embeddings through to next-token selection, and also handles MLP and CNN inference. Confirmed working models include GPT-2, Gemma 3, and Llama 2.
Adoption suits projects requiring cryptographic verification of LLM inference at scale, particularly where proof generation speed matters. The tool preserves model accuracy at 12-bit quantization and supports horizontal proof distribution and GPU acceleration. The repository is structured as a Rust workspace with the zkml crate as the core proving library, complemented by a client stack for job submission, a tensor storage layer supporting local and S3-compatible backends, and debugging tools. Developers should start with the zkml README for installation, model setup, and the end-to-end bench-llm tutorial.
The project shows active development with substantive work across multiple specialized crates addressing different layers of the proving pipeline. The codebase demonstrates careful attention to performance optimization, as evidenced by the detailed benchmark methodology and throughput measurements. Development activity spans infrastructure concerns like telemetry and storage abstraction alongside core cryptographic proving logic, indicating a mature approach to building a production-grade system rather than a research prototype.
