deep-prove

Why we included this project

When a verifier has to trust that a given model output really came from that model, DeepProve offers a practical route: it turns the neural network's forward pass into a zero-knowledge proof, so anyone can confirm the inference was computed correctly without re-running the model. That is the kind of guarantee you want for regulated reporting, audit trails, or on-chain verifiable inference, where a compact proof that a third party can check in a couple of seconds beats shipping the whole model. The Rust workspace covers the full pipeline, from quantizing models and proving each transformer layer to running a worker that serves proof-generation jobs, and the published benchmarks show end-to-end proving for GPT-2, Gemma 3, and Llama 2. Expect to budget real compute: proving is heavier than ordinary inference, so this suits teams already comfortable with zk tooling and willing to run a high-memory CPU server or GPU cluster.