EZKL

EZKL is an open toolkit for generating zero-knowledge proofs of machine-learning inference. Apache 2.0. The workflow: convert an ONNX model into a ZK circuit, generate a proof that a specific input produced a specific output under that model, let anyone verify the proof without seeing the model weights or the input. This makes private inference auditable and lets provers attest to model outputs without revealing inputs. EZKL matters because it is one of the leading bets on cryptographic, silicon-agnostic verifiable inference. Compared to siblings: Modulus Labs (similar mission, different cryptographic backend), Lagrange (ZK coprocessor with the DeepProve product line, more recent and well-funded), and the hardware-attestation alternative (TEEs like NVIDIA H100 Confidential Computing, Apple PCC, Phala). The trade is structural: hardware-attestation is fast but ties trust to the silicon vendor; ZKML is silicon-agnostic but slow. Production-readiness: beta for small models, research-grade for anything approaching frontier scale. The honest performance ceiling: ZK-proven inference of a 7B model is currently 100x+ slower than plaintext inference. Used in production at small scale by specific verifiable-AI applications (model marketplaces, auditable inference services). The strategic question for ZKML: do the cryptography optimizations close the perf gap enough to make this practical at frontier scale, or does hardware-attested confidential compute win the verifiable-AI conversation by default.