obelyzk

Python SDK for ObelyZK -- verifiable ML inference on Starknet.

All proofs use full OODS + Merkle + FRI + PoW (trustless) verification on Starknet Sepolia.

Installation

pip install obelyzk

Quick Start

from obelyzk import ObelyzkClient

client = ObelyzkClient()

# Prove and verify on-chain
result = client.prove(
    model="smollm2-135m",
    input=[1.0, 2.0, 3.0],
    on_chain=True,
)

print(f"Output: {result.output}")
print(f"TX: {result.tx_hash}")
print(f"Verified: {result.verified}")

API Reference

ObelyzkClient(url?, api_key?)

# Hosted prover (default)
client = ObelyzkClient()

# Custom prover
client = ObelyzkClient(url="http://your-gpu:8080")

# With API key
client = ObelyzkClient(api_key="your-key")

client.prove(model, input, on_chain?, recursive?)

Prove a model execution.

result = client.prove(
    model="smollm2-135m",       # model name or HuggingFace ID
    input=[1.0, 2.0, 3.0],     # input tensor
    on_chain=True,               # submit to Starknet (default: False)
    recursive=True,              # use recursive STARK (default: True)
)

# result.output          -> list[float]   model output
# result.proof_hash      -> str           Poseidon hash
# result.tx_hash         -> str | None    Starknet TX (if on_chain)
# result.verified        -> bool | None   on-chain status
# result.prove_time      -> float         seconds (~102s for SmolLM2)
# result.recursive_time  -> float         seconds (~3.55s)
# result.felts           -> int           calldata size (~942)
# result.model_id        -> str           hex identifier

client.attest(model, input, submit_on_chain?)

Full attestation with on-chain submission.

attestation = client.attest(
    model="smollm2-135m",
    input=[1.0, 2.0, 3.0],
    submit_on_chain=True,
)

client.models()

List available models.

models = client.models()
for m in models:
    print(f"{m.name}: {m.params} params, {m.layers} layers")

client.job(job_id)

Check async job status.

job = client.job("job-abc123")
print(f"Status: {job.status}, Progress: {job.progress}%")

Async Support

import asyncio
from obelyzk import AsyncObelyzkClient

async def main():
    client = AsyncObelyzkClient()

    # Prove with async/await
    result = await client.prove(
        model="smollm2-135m",
        input=[1.0, 2.0, 3.0],
        on_chain=True,
    )
    print(f"TX: {result.tx_hash}")
    print(f"Verified: {result.verified}")

    # List models
    models = await client.models()
    for m in models:
        print(f"{m.name}: {m.params}")

    # Async attestation
    attestation = await client.attest(
        model="smollm2-135m",
        input=[1.0, 2.0, 3.0],
        submit_on_chain=True,
    )

asyncio.run(main())

Supported Models

On-Chain Verification

Proofs are verified on the ObelyZK Recursive Verifier contract using full OODS + Merkle + FRI + PoW (trustless):

• Contract: 0x707819dea6210ab58b358151419a604ffdb16809b568bf6f8933067c2a28715
• Network: Starknet Sepolia
• Verification: Full OODS + Merkle + FRI + PoW (trustless)
• Felts: ~942 per proof (49x compression)
• Cost: ~$0.02 per verification

Verify independently:

from starknet_py.net.full_node_client import FullNodeClient

node_client = FullNodeClient(
    node_url="https://starknet-sepolia.g.alchemy.com/starknet/version/rpc/v0_8/demo"
)

result = await node_client.call_contract(
    contract_address=0x707819dea6210ab58b358151419a604ffdb16809b568bf6f8933067c2a28715,
    entry_point_selector="get_recursive_verification_count",
    calldata=[model_id],
)

print(f"Verification count: {result[0]}")

Environment Variables

Self-Hosted Prover

client = ObelyzkClient(url="http://your-gpu:8080")

See the Self-Hosting Guide for GPU setup.

Links

• GitHub
• PyPI
• TypeScript SDK
• CLI