AI Potluck
Back to Gap Map Product / UX / Assurance & compliance evidence

DeepProve

Lagrange Labs
open source / Overall score: 3.8

DeepProve is a zero-knowledge proof system for neural-network inference that proves full transformer forward passes, with GPT-2, Gemma 3 and Llama 2 confirmed end to end, as well as MLP and CNN models. It uses sumchecks and logup GKR rather than arithmetic circuits to keep proving time sublinear in model size. Lagrange Labs publishes it as a Rust workspace.

Tagged independently verifiable. The repository label reports Apache-2.0, but the LICENSE body is Lagrange's own evaluation-only agreement.

Openness

5 medium confidence
5.0
license
Apache-2.0/MIT(OSI
source
public(Rust workspace: the zkml proving crate plus the deep-prove, utils, tenstore, tenvis and telemetry crates)
core features withheld
no — the proving library and the worker and CLI that run it locally are all in the published workspace

The workspace manifest declares MIT OR Apache-2.0, and as ruled that package metadata governs. The LICENSE file and the README contradict it: the LICENSE is a revocable Lagrange grant for internal testing and evaluation, or unmodified use with Lagrange's platform, and the README says the code is licensed under the Lagrange License. Two of three first-party documents disagree with the reading, hence medium confidence. The whole prover, worker and CLI are in the published workspace.

Adoption

2 low confidence
2.0

GitHub stars are the only signal: DeepProve publishes no package, and the zkml crate of that name on crates.io belongs to another project. A star is not a use.

Capability

5 medium confidence
5.0

DeepProve produces a proof that a given model computed a given output, checkable without trusting whoever ran it, and it reaches full LLM forward passes. That matches NVIDIA's hardware attestation as the strongest check the category offers.

Verified 2026-09-26