📘 ZK-DeepSeek Model Card

A Verifiable Large Language Model with Zero-Knowledge Proofs

For more details, please refer to our paper:
📄 Zero-Knowledge Proof Based Verifiable Inference of Models
https://arxiv.org/pdf/2511.19902

---

Overview

ZK-DeepSeek is a verifiable large language model (Verifiable LLM) whose inference can be cryptographically proven using zero-knowledge proofs (zkSNARKs).

For every inference step, the model produces a succinct mathematical proof that verifies:

• The output is computed faithfully by the intended model, and
• No model parameters or internal states are revealed.

This enables trustless, provable, and privacy-preserving AI inference suitable for high-stakes environments such as blockchain, Web3, finance, governance, and distributed systems.

This repository provides:

• A fully arithmeticized version of the model (Int64-based)
• Layer-wise model loading and computing
• Per-component zero-knowledge circuits (GeMM, RMSNorm, RoPE, Softmax, SiLU, MLA, MoE)
• A recursive SNARK pipeline

---

Key Features

Verifiable Inference

Every model output is accompanied by a zkSNARK proof guaranteeing correctness.

Zero-Knowledge Privacy

The prover demonstrates correct computation without exposing any model parameters.

Full Arithmeticization

All DeepSeek-like Transformer operations are converted into constraint-friendly circuits:

• GEMM (matrix multiplication)
• RMSNorm
• RoPE
• Softmax
• Sigmoid / SiLU
• MLA (Multi-Head Latent Attention)
• MoE (Mixture-of-Experts routing)

Recursive Proof Composition

Tens of thousands of component proofs are folded into a single constant-sized proof.

---

Model Details

Item	Description
Architecture	DeepSeek-V3 style Transformer with MoE + MLA
Parameters	~671B (quantized)
Quantization	Int64 / Int32 — exact arithmetic for zk circuits
Disk Size	~2.5 TB (expanded integer representation)
Intended Use	Research & verifiable inference

Environment Requirements

Hardware

Recommended:

• NVIDIA RTX 4090 / 5090 (preferred) * 1
• 64 GB RAM
• 6 TB NVMe SSD (800 G for virtual memory)

Software

• Node.js 24.8 + TypeScript
• CUDA 12.9+
• Python 3.10
• o1js 2.10

---

Quick Start

Generate Zero-Knowledge Witnesses

Clone the repo:

git clone https://huggingface.co/arcstar-lab/ZK-DeepSeek
cd ZK-DeepSeek/inference

Install dependencies:

pip install -r requirements.txt

Compile CUDA kernels (-arch depends on your GPU):

nvcc -O3 -Xcompiler -fPIC -shared -o libint64gemm.so int64_gemm.cu -arch=sm_120

Download DeepSeek-V3 model weights and place them in:

/path/to/DeepSeek-V3

Convert weights to integer representation:

python3 ./convert2.py --hf-ckpt-path /path/to/DeepSeek-V3 --save-path /path/to/ZK-DeepSeek-Demo1 --n-experts 256 --model-parallel 1

This also stores embedding data in the zkdata folder.

Run chat interface:

./runLLM.sh
<Input your message>

The inference log will be available in the logs folder, and intermediate states in zkdata.

Generate Zero-Knowledge Proofs

Move to zk folder and install ZK dependencies:

cd ../zk
npm install
npm run build

Start proof generation for vocabulary embeddings.

python3 runZK.py

To enable additional components, uncomment lines in runZK.py such as:

    # await taskExpertSelector_gate(0, 4)
    await taskEmbed()
    # await taskAttnNorm('attn_norm', 0, 0)
    # await taskAttnNorm('q_norm')
    # await taskRope_pe()
    # await taskSoftmax('scores')
    # await taskSigmoid('gate')
    # await taskExpertSelector('gate', 0, 3)
    # await taskGemm('wkv_a1', 0, 0, 7168, 512, 112)

---

Safety & Limitations

Security Guarantees

• Proofs ensure inference correctness
• Model parameters remain private
• Inference cannot be forged or shortcut
• Compatible with on-chain trustless environments

Limitations

• Proving is computationally heavy but can be accelerated by GPU

---

License

MIT

---

Contact

Author: Edward Wang Email: [email protected]

Our next priority is to develop a GPU-accelerated version of ZK-DeepSeek, which we anticipate will yield performance improvements by several orders of magnitude. We are actively seeking funding and collaborators to help advance this line of research. If you are interested in supporting or partnering with us, we warmly invite you to get in touch by email.