#lagrange Labs is excited to unveil DeepProve-1, the first production-ready zkML system capable of generating a cryptographic proof for a complete large language model (LLM) inference. What was once considered a distant ambition verifiable AI is now a tangible reality.
With DeepProve-1, we have successfully produced a zero-knowledge proof for the full inference of OpenAI’s GPT-2, marking a major leap forward in verifiable machine learning. Beyond the technical milestone, this achievement lays the groundwork for proof systems across the next wave of LLMs such as LLAMA, Gemma, and others. Because GPT-2 and Meta’s LLAMA share key architectural patterns, DeepProve is now positioned closer than ever to proving LLAMA-scale models. Lagrange plans to bridge this final step in the months ahead, unlocking verifiability for some of the world’s most widely used open-source LLMs.
Why DeepProve-1 Matters
DeepProve-1 is an inflection point in the evolution of machine learning bringing verifiability to transformer-based AI systems for the first time. Until now, zero-knowledge proofs were limited to simpler models such as MLPs and CNNs. With this release, verifiability extends to state of the art transformer architectures, embedding cryptographic integrity into the core of modern AI.
As artificial intelligence increasingly shapes decision-making in critical domains like defense, healthcare, finance, and infrastructure, DeepProve-1 ensures that these systems can operate with mathematical guarantees of correctness and trust.
How We Proved GPT-2
Delivering verifiable inference for GPT-2 required deep collaboration across cryptography, systems engineering, and machine learning research. Since our previous milestone fully provable inferences for MLPs and CNNs our team has been laser-focused on enabling the structural and computational primitives that define transformer models.
This effort demanded major upgrades to the DeepProve framework, including:
Extending support to transformer layers and attention mechanisms.
Optimizing proof generation to handle LLM-scale computations efficiently.
Engineering novel cryptographic techniques to ensure scalability without compromising security.
DeepProve-1 is more than a proof-of-concept, it is a production-grade foundation for the zkML era of AI, where outputs are not only powerful but also provably correct.
