Framework to prove inference of ML models blazingly fast
3,353
stars
1,219
commits
Rust
primary language
May 31, 2026
updated
README
DeepProve
Zero-knowledge proof system for neural network inference, with first-class support for end-to-end LLM proving.
π Looking to run DeepProve? Start with zkml/README.md
That's where the installation steps, model setup, GPU build, and the full end-to-end bench-llm tutorial live. The rest of this page is a high-level summary of what DeepProve is and what to expect.
Overview
DeepProve is the first end-to-end zero-knowledge proof system for full LLM inference. It generates cryptographic proofs of neural network forward passes using sumchecks and logup GKR, achieving sublinear proving time in model size β orders of magnitude faster than circuit-based approaches.
Confirmed working models: GPT-2, Gemma 3, Llama 2 β all transformer layers proven end-to-end, from token embeddings through to next-token argmax. MLP and CNN inference is also supported.
This repository is a Rust workspace. The zkml crate is the core proving library; the remaining crates provide the client stack, storage layer, and developer tooling.
Headline Numbers
Single-machine inference proving on a 24-core / 504 GB CPU server:
| Model | Sequence | Prove time | Verify | Proof size | Throughput |
|---|---|---|---|---|---|
| GPT-2 | 512 tokens | 7.6 min | 1.3 s | 10.7 MiB | 1.12 tokens/s (67 tokens/min) |
| Gemma 3 | 512 tokens | 19 min | 4.3 s | 27 MiB | 0.45 tokens/s (27 tokens/min) |
- 10β30Γ faster than the previous published state of the art (e.g. zkGPT reports β 0.05 tokens/s on similar hardware).
- Accuracy preserved: β₯99.6% cosine similarity to the floating-point baseline at 12-bit quantization (GPT-2).
- Scales out: horizontal proof distribution and GPU acceleration are supported today; clusters of GPU workers are on the roadmap.
For the full methodology and a deeper benchmark sweep across sequence lengths and models, see the DeepProve paper (link to be added) and zkml/README.md.
Repository Structure
| Crate | Description |
|---|---|
zkml | Core proving library β model quantization, layer implementations (MLP, CNN, transformer), and ZK proof generation/verification |
deep-prove | Client stack β deep-prove-worker runs a proof generation server; deep-prove-cli submits proving jobs locally or to a remote proving network |
tenstore | Storage facade for persisting and retrieving tensor data; supports local and remote (S3-compatible) backends |
tenvis | Interactive CLI tool for inspecting and debugging proof data stored in tenstore |
telemetry | Shared OpenTelemetry tracing and logging setup used across all crates |
utils | Shared utility helpers: CSV recording, memory tracking, statistical summaries |
Licensing
Licensed under the Lagrange License.
Acknowledgements
This project builds upon the work from scroll-tech/ceno, reusing the sumcheck and GKR implementation from that codebase.
Contributors
(top 30 of 37)
Similar repos
Languages
Rust
90.2%
Python
9.6%
Framework to prove inference of ML models blazingly fast
3,353
stars
1,219
commits
Rust
primary language
May 31, 2026
updated
README
DeepProve
Zero-knowledge proof system for neural network inference, with first-class support for end-to-end LLM proving.
π Looking to run DeepProve? Start with zkml/README.md
That's where the installation steps, model setup, GPU build, and the full end-to-end bench-llm tutorial live. The rest of this page is a high-level summary of what DeepProve is and what to expect.
Overview
DeepProve is the first end-to-end zero-knowledge proof system for full LLM inference. It generates cryptographic proofs of neural network forward passes using sumchecks and logup GKR, achieving sublinear proving time in model size β orders of magnitude faster than circuit-based approaches.
Confirmed working models: GPT-2, Gemma 3, Llama 2 β all transformer layers proven end-to-end, from token embeddings through to next-token argmax. MLP and CNN inference is also supported.
This repository is a Rust workspace. The zkml crate is the core proving library; the remaining crates provide the client stack, storage layer, and developer tooling.
Headline Numbers
Single-machine inference proving on a 24-core / 504 GB CPU server:
| Model | Sequence | Prove time | Verify | Proof size | Throughput |
|---|---|---|---|---|---|
| GPT-2 | 512 tokens | 7.6 min | 1.3 s | 10.7 MiB | 1.12 tokens/s (67 tokens/min) |
| Gemma 3 | 512 tokens | 19 min | 4.3 s | 27 MiB | 0.45 tokens/s (27 tokens/min) |
- 10β30Γ faster than the previous published state of the art (e.g. zkGPT reports β 0.05 tokens/s on similar hardware).
- Accuracy preserved: β₯99.6% cosine similarity to the floating-point baseline at 12-bit quantization (GPT-2).
- Scales out: horizontal proof distribution and GPU acceleration are supported today; clusters of GPU workers are on the roadmap.
For the full methodology and a deeper benchmark sweep across sequence lengths and models, see the DeepProve paper (link to be added) and zkml/README.md.
Repository Structure
| Crate | Description |
|---|---|
zkml | Core proving library β model quantization, layer implementations (MLP, CNN, transformer), and ZK proof generation/verification |
deep-prove | Client stack β deep-prove-worker runs a proof generation server; deep-prove-cli submits proving jobs locally or to a remote proving network |
tenstore | Storage facade for persisting and retrieving tensor data; supports local and remote (S3-compatible) backends |
tenvis | Interactive CLI tool for inspecting and debugging proof data stored in tenstore |
telemetry | Shared OpenTelemetry tracing and logging setup used across all crates |
utils | Shared utility helpers: CSV recording, memory tracking, statistical summaries |
Licensing
Licensed under the Lagrange License.
Acknowledgements
This project builds upon the work from scroll-tech/ceno, reusing the sumcheck and GKR implementation from that codebase.
Similar repos
Contributors
(top 30 of 37)
Languages
Rust
90.2%
Python
9.6%