DeepProve-1 proves GPT-2 inference in zero-knowledge, setting the stage for verifiable AI across future large language models.
Lagrange Labs has announced DeepProve-1, the first production-ready zkML system capable of generating a cryptographic proof for the full inference of a large language model.
The achievement establishes a foundation for supporting models such as LLAMA and Gemma, with Lagrange aiming to extend DeepProve-1’s reach to widely adopted open-source LLMs in the coming months.
DeepProve-1 introduces verifiability as a feature for modern AI systems, extending beyond earlier support for simpler models to full transformer architectures. This milestone comes as AI increasingly powers critical decision-making in sectors including defence, healthcare, finance, and infrastructure.
Proving GPT-2 required advances in cryptography, systems engineering, and machine learning. Lagrange’s team expanded the DeepProve framework to support complex graph structures, transformer-specific layers such as softmax and multi-head attention, and widely used model formats like GGUF.
The company has prioritised future performance improvements, with ongoing work in cryptographic optimisation and parallel computation to ensure practical deployment in real-world settings.
DeepProve-1 demonstrates the feasibility of zero-knowledge proofs for large language models, bringing Lagrange closer to proving models such as LLAMA and Mistral.
Stay updated on the crypto space by following our socials
