DESILO’s THOR framework runs open source LLMs fully under homomorphic encryption without retraining, delivering near practical speed and redefining privacy-safe AI deployment.
DESILO Inc. has unveiled THOR, a breakthrough framework that enables large language model (LLM) inference to run fully under homomorphic encryption. The innovation allows widely used open source LLMs to operate without retraining, keeping both inputs and outputs encrypted, marking a significant advance in privacy-safe AI.
The company’s joint research with Professor Miran Kim’s team at Hanyang University has been accepted for presentation at ACM CCS 2025, one of the world’s leading peer-reviewed security conferences alongside IEEE S&P and USENIX Security.
THOR achieves near practical runtime, processing roughly 128 tokens (two sentences) on a single GPU at deployment-relevant speed. Core matrix multiplication performance improved 5.3× for plaintext to ciphertext operations and 9.7× for ciphertext to ciphertext operations, setting a new benchmark for privacy-preserving AI. According to the team, THOR is the first known framework to run an entire existing LLM under homomorphic encryption without retraining while maintaining near practical performance.
Seungmyung Lee, CEO of DESILO, said: “This CCS acceptance recognises a breakthrough in running large language models fully under homomorphic encryption, marking an important academic milestone in privacy-preserving AI research. It also highlights our dual focus: driving forward homomorphic encryption research with partners like Cornami, and accelerating product development to bring trusted Privacy AI into real-world applications.”
THOR forms the technical foundation for upcoming DESILO solutions such as Harvest™, designed to enable secure, privacy-safe analysis across multiple institutions, setting a new benchmark for open source AI frameworks in privacy-focused research and real-world deployment.
