Lagrange Labs announced at Google’s Zero-Knowledge event in Singapore that its DeepProve platform has successfully generated the world’s first cryptographic proof for Google’s Gemma3 large language model inference. This development demonstrates the capability of advanced AI architectures to integrate cryptographic guarantees of correctness without compromising performance or innovation.
This achievement is seen as a significant inflection point for both AI and zero-knowledge (ZK) technology, indicating that ZK solutions can adapt in real-time to sophisticated model architectures emerging from leading AI research. Ismael Hishon-Rezaizadeh, CEO and co-founder of Lagrange Labs, stated, “Gemma3 represents a clear departure from older GPT-style models with smarter attention mechanisms and novel optimizations. By proving Gemma3, we’re demonstrating that verifiable AI isn’t a retrofit for yesterday’s models. It’s living infrastructure that evolves with the frontier of AI research.”
Gemma3’s advanced architecture presented several cryptographic challenges that necessitated fundamental innovations within DeepProve’s proving system. These included adapting to Group Query Attention (GQA), which shares Key and Value tensors across multiple Query heads, and implementing new proof structures for its asymmetric attention patterns. DeepProve also developed flexible attention masking proofs for Gemma3’s alternating local sliding-window and global attention layers. Furthermore, the system integrated proofs for Rotary Positional Encoding (RoPE) for efficient scaling and built optimized proof circuits for the architecture’s dense RMSNorm layers, which occur five to six times per attention block.
The successful proving of Gemma3 aims to eliminate the need for organizations to choose between deploying cutting-edge models and adhering to regulatory requirements. This capability could enable healthcare systems to verify diagnostic models and ensure correct patient data processing, allow defense contractors to confirm mission-critical AI operates within defined parameters, and empower financial institutions to provide cryptographic receipts for AI-driven decisions. Hishon-Rezaizadeh added, “Organizations no longer need to choose between innovation and verifiability. DeepProve on Gemma3 proves that compliance keeps pace with model upgrades.”
DeepProve’s Gemma3 integration leverages Lagrange’s established production record, which includes over 11 million zero-knowledge proofs generated and 3 million AI inferences verified. The system maintains DeepProve’s reported 158x performance advantage over competing zkML solutions while supporting advanced model architectures. Hishon-Rezaizadeh concluded, “With DeepProve on Gemma3, proving AI correctness isn’t hypothetical anymore. It’s cryptographically guaranteed.”
Lagrange Labs operates a universal ZK Prover Network on EigenLayer, offering cryptographic verification for AI, rollups, and cross-chain applications. The company has formed partnerships with entities such as NVIDIA, Intel, and AWS, serving enterprises that require verifiable computation at scale. Lagrange Labs has secured $21.5 million in funding from investors including Founders Fund and 1kx, and plans to launch its LAG token in July 2025. DeepProve is described as a fast, production-ready zkML system designed for cryptographic verification of AI model inferences, supporting transformer architectures like GPT-2, LLAMA, and now Gemma3, to provide mathematical guarantees of AI correctness.
