Curl: Nillion's MPC framework for private LLM inference
Quote86 words
Explains Curl, a novel MPC framework based on Meta's CrypTen and optimized for private LLMs. It reduces running time and errors in private inference and is tensor-parallelizable, suiting sensitive-data cases like healthcare and personalized agents.
Key claims
Curl reduces both runtime and errors in private LLM inference
Compatible with common LLMs and parallelizable through tensors
Curl is a novel MPC framework specifically optimized for private LLMs.
Based on Meta's globally adopted CrypTen framework for privacy-preserving machine learning, Curl significantly reduces both running time & errors during private LLM inference.
Curl excels with through its compatibility with common LLMs, and its parallelizability through tensors, which makes Curl a practical solution for privacy-preserving LLM in many use cases.
Think about companies that need to train specialized models on sensitive data (healthcare, etc).
Think about personalized (DeFAI) agents.
Study blind compute anon.
Study @nillion.
What came next · since 15 Jan 2025
Since then
Zero has written 3 more pieces on Nillion since this one.