---
title: Encrypted AI Inference Now Has a Benchmark Anyone Can Rerun
description: The FHE Benchmarking Suite has added a BERT workload. DESILO's THOR is the reference implementation, running on encrypted data in about two minutes.
author: Darie Nani (Editor-in-Chief)
date: 2026-07-29T14:30:00.000Z
updated: 2026-07-29T14:56:35.330Z
canonical: https://www.sovereignmagazine.com/article/encrypted-ai-inference-fhe-benchmark-thor
image: https://cdn.nanimediahouse.com/pexels-a-woman-with-digital-code-projections-on-her-face-representi-3861969.jpg
categories: Science & Tech
content_type: News
region: Global
publication: Sovereign Magazine
access: members
schema_type: Article
---

Anyone can now download the code, run a language model over data that is never decrypted, and compare the result against a published reference. The FHE Benchmarking Suite maintained by HomomorphicEncryption.org has added a BERT inference workload, and its reference submission follows a method called THOR, developed by the Korean company DESILO with researchers at Hanyang University, Yonsei University and the University of Texas Health Science Center at Houston.

Fully homomorphic encryption lets a computer calculate directly on encrypted data without decrypting it first, so a cloud service can process information it cannot read. The obstacle has never been whether it works. It has been what it costs, and until now every team reported that cost on its own hardware under its own settings.

## The Workload Is BERT-Base on a Sentence-Pair Task

---

*This article is only available to registered readers. Visit the article URL to read the full piece.*