— Establishing a universal benchmark for Private AI and encrypted Large Language Model (LLM) inference
SEOUL, South Korea, July 29, 2026 /PRNewswire/ — 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.
“Homomorphic encryption has advanced quickly, but because measurement conditions differ from one result to the next, systematic, objective comparison of real progress has remained out of reach. Much like MLPerf in machine learning, the FHE Benchmarking Suite serves as a reference point shared across the entire community,” said Shruthi Gorantala of Google, a co-lead of the FHE benchmarking suite. “With Transformer inference now included in the benchmark, encrypted inference is no longer a one-off demonstration but a result anyone can reproduce, compare, and build on. That is exactly how an ecosystem is formed.”
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.
“THOR’s inclusion in the FHE Benchmarking Suite reflects our belief that this field should advance in the open, on shared standards; achievements must be verified, not merely asserted,” said Seungmyung Lee, CEO of DESILO. “We are proud to present the first implementation of a common foundation for encrypted language-model inference and to release it as a common standard, and we will continue working to carry fully homomorphic encryption from the research stage into real-world Private AI.”
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 Standardization (ISO).
About DESILO
DESILO Inc. is a leading deep-tech company dedicated to resolving the complex data-privacy bottlenecks that arise in AI environments. Specializing in fully homomorphic encryption (FHE) and advanced cryptographic frameworks, DESILO builds the core infrastructure for a future of highly secure Private AI, enabling the global technology ecosystem to draw full value from sensitive data without compromising security. DESILO is backed by leading investors including NAVER D2SF, Bon Angels Venture Partners, KB Investment, Schmidt, SV Investment, the Korea Development Bank, and LG Electronics.
ข่าวที่เกี่ยวข้อง
- Electrolux Group Interim report Q2 2026
- Dinto Solar’s HJT Modules Gain India BIS CRS Registration Amid Rising Demand for Higher-Value Solar Solutions
- JustMarkets Analyzes the Core Drivers Behind 2026 Gold Price Movements
- eight Telecom marks annual eight Day celebrations with new eight Shop and $88,888 lucky draw
