Zero Knowledge · Zero Knowledge Podcast

ZKTorch & the Evolution of ZKML with Daniel Kang

July 30, 2025·51 min·4 clips

Daniel Kang returns to update on ZKML's progress and his new project ZK-Torch, bridging speed and flexibility.

ZKML is not one tidy lane anymore. Daniel Kang returns for a catch-up on how the tools, tradeoffs, and expectations have moved since early 2023. Anna keeps the conversation practical, which helps. Instead of treating zero knowledge plus machine learning as a shiny pairing, she keeps asking what the tools prove and what a developer would actually do with them. The first split is flexibility. Daniel compares ZK-LLM and ZK-CNN with broader approaches like Ezekiel and ZKML, with speed and compute sitting on the other side of the tradeoff. A specialized system can be faster, but it cannot handle every model shape or every application a team may care about. The more generic route sounds easier to live with. Give it a model specification, maybe GPT-4O or a classical ResNet, and the system translates that into a circuit before running the proving underneath. Anna asks the useful, slightly boring question: does the model developer rebuild the whole thing by hand, or can they plug in the specification and let the proving system handle the rest? Daniel says it is closer to plug in. That answer makes ZKML feel less like custom cryptography work and more like tooling that could sit near ordinary ML workflows. Then the episode gets into scope. Prompts, inputs, and outputs all come up, along with whether these systems are broad proof machines or narrower tools. The answer depends on the project. Ezekiel and ZKML can span those uses, while ZK-LLM is built for the LLM-specific case and does not cover classification models. ZK-Torch is the newer piece: Daniel frames it as a general-purpose tool for proving things about ML systems quickly while keeping up with emerging models. The ending stays open-source and practical. The license is either MIT or Apache, Daniel wants people to build on it, and he wants to see a thousand flowers bloom.

As heard by us

A practical tour of ZKML tradeoffs, tooling, and ZK-Torch's aims.

Daniel Kang gives a clear update on where ZKML has gone since early 2023. The episode works best when it stays with the concrete tradeoffs. It moves through ZK-CNN, ZK-LLM, Ezekiel, ZKML, and ZK-Torch, and it does a good job of showing both what these systems are supposed to…

Read the full review in PlayNext →

Why you'd press play

You want the clearest current map of ZKML, with ZK-Torch as the newest waypoint.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Zero Knowledge