DESILO's THOR Selected as Reference Implementation for Encrypted AI in Global FHE Benchmarking Suite

DESILO Inc. today announced that THOR — the world's first framework to run the entire LLM inference process directly on encrypted data — has been officially added to the global FHE Benchmarking Suite as the reference implementation for encrypted language-model inference.

Through this, a common standard on which anyone can reproduce and compare the performance of encrypted AI inference has been established, opening the door for enterprise in healthcare, finance, and the public sector to safely deploy powerful AI models on sensitive data without risk of exposure.

Fully homomorphic encryption (FHE) is a technology that enables model computation directly on encrypted data without decryption, providing the strongest level of privacy protection in AI environments. Yet while technology has advanced dramatically, results have been reported individually under differing hardware and parameter conditions, making real progress difficult to assess objectively.

The FHE Benchmarking Suite is intended to resolve this by defining shared, reproducible workloads and performance measures for the community. Within it, THOR is included as the first reference implementation for encrypted language-model inference, serving as the reference point against which related technologies can be compared and verified going forward.

THOR runs the widely used AI language model BERT with the data kept encrypted on a single GPU, completing the benchmark inference workload while maintaining accuracy on par with the unencrypted model (within about one percentage point). It processes an existing open-source AI model end to end — from input through computation to output — entirely on encrypted data, without any separate retraining. Running an entire LLM under encryption at a practical speed in this way is a world first for THOR.

Performance highlights include:

Drastic Speed Improvement: Inference time has been cut from about 10 minutes to about 2 minutes through continued optimization.

Matrix Acceleration: Accelerates matrix multiplication — the core of the computation — by up to 9.7 X over the previous state of the art, markedly improving processing speed long regarded as FHE's greatest challenge.

Universal Standard: THOR's implementation and performance-measurement methodology in the suite can be reproduced and verified by anyone under the same conditions, serving as the baseline for comparing encrypted language-model inference technologies going forward.

DESILO continues to advance FHE-based Private AI for enterprise and public-sector use and participates in fully homomorphic encryption standardization discussions at both the U.S. National Institute of Standards and Technology (NIST) and the International Organization for S...