Built on STWOStarknet Mainnet

Prove Any AI Model.
Trust No Intermediary.

ObelyZK verifies neural network inference entirely on-chain — from a single matrix multiply to a 14-billion parameter transformer. GPU-accelerated zero-knowledge proofs mean anyone can confirm an AI produced a specific output, without re-running the model or trusting a third party.

0
Lines of Code
Rust + Cairo
0
Tests Passing
979 lib + 88 integ + 385 Cairo
0
CUDA Kernels
Sumcheck · GEMM · LogUp
0.00s
Per-Layer Prove
Qwen3-14B · GPU

Why GPU + SIMD for DePIN.

The M31 prime enables native 32-bit SIMD on every CPU. Our CUDA kernels unlock GPU-class throughput. Together, any DePIN node becomes a prover.

M31 = Native 32-bit

The Mersenne-31 prime (2³¹ − 1) fits a single CPU SIMD lane. Every x86/ARM chip does M31 arithmetic at near-native speed — perfect for DePIN edge nodes.

17 CUDA Kernels

Custom sumcheck, GEMM, MLE restrict, and LogUp kernels. H100/A100/4090 nodes get 2.56x+ speedup on MLE operations alone.

Hybrid Execution

GPU for Merkle/FRI/quotients above thresholds, CPU SIMD for FFT/interpolation. Optimal work distribution — uses every core and every GPU simultaneously.

Multi-GPU Scaling

DeviceGuard RAII + propagate_device() across Rayon workers. Scales horizontally across GPU-rich DePIN providers. Any node with a GPU can earn.

17 CUDA Kernels — Compiled at runtime via cudarc + NVRTC
Sumcheck
sumcheck_round_kernelsumcheck_reduce_kernelmle_fold_kernel
Forward Pass
m31_gemv_kernelm31_gemm_kernelm31_add_kernelm31_mul_kernelm31_relu_kernelm31_layernorm_kernelm31_rmsnorm_kernel
MLE Restrict
m31_restrict_rows_kernelm31_restrict_cols_kernel
LogUp STARK
logup_denominator_kernellogup_3way_round_kernellogup_4way_reduce_kernellogup_3way_fold_kernelcombine_blocks_kernel

We dominate every axis.

No other ZKML framework combines GPU acceleration, full transformer support, on-chain verification, and a privacy protocol.

FeatureObelyZKGizaEZKLLagrangezkPyTorch
Proving SystemSTWO Circle STARKSTWO Circle STARKHalo2 SNARKGKRExpander
GPU Accel.17 CUDA kernelsNone (Phase 3)Yes (Halo2)Yes (distributed)Yes
PrivacyFull VM31NonePartialNoneNone
LLM ScaleQwen3-14B (40L)Phase 1 onlyNo LLM supportGPT-2 / Gemma3Llama-3 (150s/tok)
StarknetMainnetNone (Phase 3)NoNoNo
ML Operators15 components11 primitives~20 constraints~10 opsLimited
TransformerFull (Attn+RoPE+GQA)PartialNoPartialPartial
Codebase127,500 lines~5,000 lines~50,000 linesClosed source~10,000 lines
Tests1,452< 100~500UnknownUnknown
vs LuminAIR
12-18 months ahead

LuminAIR is still in Phase 1 implementing basic operators. We're proving 14B-parameter transformers with GPU acceleration, privacy, and on-chain verification.

vs Lagrange
112x larger models

Lagrange proved GPT-2 (124M params, a 2019 model). We proved Qwen3-14B — a 2025-era model that is 112x larger. No privacy. No Starknet.

vs EZKL
STARKs beat SNARKs

EZKL uses Halo2 SNARKs — larger proofs, more expensive verification, no Starknet integration. They cannot handle LLM-scale models.

15 provable ML components.

Every operation in a modern transformer — from matmul sumcheck to multi-head attention to rotary embeddings — has a dedicated proof component.

OperationProof Method
MatMulGKR Sumcheck
Multi-Head Attention5-stage composed proof
RoPELogUp STARK
ReLU / GELU / SigmoidLogUp STARK
SoftmaxLogUp STARK
LayerNormLogUp + variance mask
RMSNormLogUp + variance mask
EmbeddingTable lookup
Conv2Dim2col + MatMul sumcheck
Quantize / DequantizeLogUp range check
Element-wise Add/MulDegree-1/2 check
Poseidon2-M31 AIRConstraint framework
Supported Architectures
Qwen3-14BLlama 2/3MistralGemmaGPT-2BERT

+ any HuggingFace transformer with standard config.json

Model Formats
ONNXtract-onnx parser
SafeTensorsmemmap2, f16/bf16/f32 + GPTQ/AWQ INT4
HuggingFaceconfig.json + multi-shard auto-detection
StreamingTile-level mmap for 100B+ models
Performance
3.04s
Qwen3-14B per layer (GKR: 0.80s)
86,723
Felts calldata for 40-layer proof

VM31 — Nobody else has this.

A complete privacy system built natively on STWO's M31 field. Poseidon2-M31 hashing, shielded notes, nullifiers, encrypted transfer — all proven with STARKs. No other ZKML framework has anything like it.

Poseidon2-M31 Hash

t=16, rate=8, d=5 — 248-bit security, native to STWO field. No external crypto dependencies.

Shielded Notes

11 M31-input commitments. Full UTXO model with Poseidon2-compress Merkle tree (depth 20).

Nullifier System

12 M31-input nullifiers for double-spend prevention. Deterministic, verifiable on-chain.

Counter-Mode Encryption

Poseidon2-M31 stream cipher. Zero external crypto deps — entire stack uses one hash function.

Transaction Circuits
Deposit

Public amount → shielded note + STARK proof

Withdraw

Shielded note → public amount + nullifier + STARK proof

Spend

2-in / 2-out private transfer + STARK proof

Batch Prove

Multi-component composition — bundle multiple txs

Privacy SDK Modules
HD WalletTX BuilderNote StorePool ClientRelayerTree Sync

19,683 lines of Cairo.

Deployed on Starknet. Complete GKR verification, shielded pool, aggregated weight binding, and ML-specific AIR constraints — all on-chain.

verifier.cairo1,611Core STARK proof verifier (9-TX multi-step)
vm31_pool.cairo1,556Shielded pool — deposit/withdraw/spend
aggregated_binding.cairo994Aggregated weight binding (RLC mode 4)
ml_air.cairo962ML-specific AIR constraints
vm31_merkle.cairo724Poseidon2-M31 Merkle tree (depth 20)
gkr.cairo546GKR sumcheck on-chain verifier
model_verifier.cairo538End-to-end model verification

Live proof streaming.

Watch proofs generate in real time. WebSocket broadcast, 3D visualization, 13 event types — from sumcheck rounds to GPU status to proof completion.

WebSocket
WsBroadcastSink

tokio::sync::broadcast fan-out. Connect any number of clients. Three.js helix + Chart.js dashboard embedded in prove-server.

Visualization
Rerun.io 3D

Rerun 0.21 integration for 3D proof topology visualization. Attention heatmaps, sumcheck descent charts, GPU metric dashboards.

Events
13 Event Types

ProofStart, LayerStart, SumcheckRound, LayerEnd, GpuStatus, ProofComplete, CircuitNodeMeta, ActivationStats, and more.

What's shipped. What's next.

Shipped
  • GPU sumcheck (17 kernels)
  • Qwen3-14B 40-layer proof
  • On-chain GKR verifier
  • VM31 privacy protocol
  • Proof streaming (WebSocket)
  • Audit pipeline (M31-native)
In Progress
  • Deploy updated Cairo verifier with RLC mode 4
  • Reduce on-chain TX count (9 → 3)
  • Confidential computing (TEE + GPU attestation)
Next
  • DePIN prover network launch
  • Cross-chain proof relay
  • Herodotus/Integrity integration
  • Sub-second proving for small models

Own the ZKML narrative.

The first ecosystem to ship production ZKML wins the AI + crypto narrative. ObelyZK is 127,500 lines of battle-tested code — the official ZKML stack for Starknet.

Upstream
Official ZKML toolkit for the Starknet ecosystem
DePIN
GPU prover network — nodes earn fees for ML verification
Optimize
Reduce on-chain cost (9 TXs → 3, lower calldata)
Integrate
Herodotus/Integrity for cross-chain proof relay
info@obelysk.xyz