DeepProve is a free, open-source zero-knowledge proof system for machine-learning inference, built by Lagrange Labs. It generates cryptographic proofs that a neural network — including a full large language model — actually ran on a given input and produced a given output, without revealing the model or the data, and does so orders of magnitude faster than earlier circuit-based approaches.
DeepProve is an open-source project on GitHub with about 3,400 stars, actively developed in Rust. It is the first end-to-end ZK proof system for complete LLM inference: transformer layers from token embeddings to the final argmax are proven using sumcheck and LogUp GKR protocols, achieving sublinear proving time in model size. GPT-2, Gemma 3 and Llama 2 are confirmed working end to end, with MLP and CNN inference also supported, at 10–30× the proving throughput of the previous state of the art while preserving at least 99.6% cosine similarity to the floating-point baseline. The workspace includes the core zkml proving library, a worker and CLI client stack, and tooling for distributed proving with GPU acceleration.
DeepProve is free and open source under the Lagrange License. You can run the prover on your own hardware; there are no fees for the software itself.
DeepProve is for researchers and protocol engineers working on verifiable AI — anyone who needs cryptographic assurance that a model’s output was genuinely produced by the claimed model on the claimed input.