Next-generation FHE computation engine
maximizing speed and usability
Fully homomorphic encryption (FHE) theoretically enables perfect privacy-preserving computation, but in practice computational speed, accuracy, and usability have always sat in a trade-off.
DESILO FHE Library optimizes all three — speed, accuracy, and usability — at the same time, making it a production-grade homomorphic encryption library accessible to everyone from FHE newcomers to expert algorithm designers.
![[ System diagram — designer to provide ] · FHE Library stack (schemes · GPU/CPU · Python wrapper · multi-party · bootstrapping)](/images/solutions/fhe-library-diagram.png)
FHE infrastructure proven in both research and production
Multi-scheme support
Supports both industry-standard RNS-CKKS and the next-generation GL Scheme (5th-generation FHE). RNS-CKKS is optimized for vector operations and the GL Scheme for large-scale matrix operations — pick the right scheme for your workload.
GPU acceleration and CPU parallelism
Built on C++/CUDA, runs on NVIDIA GPUs and falls back to parallel CPU execution where GPUs are unavailable.
Python wrapper
Includes APIs for encoding and encrypting PyTorch tensors directly, integrating naturally with existing AI workflows.
Multi-party computation
Native support for joint key generation and distributed decryption — enabling collaboration scenarios such as data clean rooms and federated analysis.
Full bootstrapping support
Supports the full range of bootstrapping variants — standard, small, lossy, and merge — so accuracy and performance can be tuned per scenario.
Best-in-class FHE performance — backed by the research ||that made it possible
The only implementation of the next-generation GL Scheme
The world's first implementation of the 5th-generation homomorphic encryption GL Scheme, co-developed by FHE inventor Craig Gentry and DESILO. Delivers dramatic performance gains over CKKS on the large-scale matrix operations at the heart of AI workloads.
Research foundation validated at top-tier venues (CCS, Crypto, etc.)
The library powering THOR — the world's first FHE LLM, presented at ACM CCS 2025 — and the implementation behind two papers accepted to Crypto 2026. Stability validated at academic-standard rigor.
Intuitive API design
Abstractions deliberately tuned so researchers aren't bogged down by cryptographic detail. Intuitive APIs like encode/encrypt/decrypt and PyTorch compatibility minimize the learning curve.
A global standard reference
Selected as the LLM reference implementation in the HES FHE Benchmarking Suite jointly led by AWS and Google — establishing the benchmark for global FHE performance measurement.
From the research lab to production services
Prototyping new FHE algorithms, optimizing bootstrapping, designing multi-party protocols, and other academic research.
Research on model inference and training over encrypted data, and the design of FHE-friendly model architectures.
Used as the core computation engine in FHE-based LLM and generative-AI inference systems, as exemplified by THOR.
Acts as the backbone of DESILO Data Clean Room (DCR), enabling secure data collaboration through joint key generation and distributed decryption.