Product for Windows 11
Albireo
Your private AI assistant, on your PC. Albireo runs language models on the computer where it is installed: chat, documents, images, agent tools and an OpenAI-compatible interface, with nothing sent outside.

Cydral Technology is a French open-source initiative. Its C++ library and its two free Windows applications, Albireo and Pseudalis, run language models and protect personal data directly on your own machine: no account, no subscription, no cloud.
Everything is written in C++, from the tensor operations to the Windows installer. The library provides the engine; the products bring it to every desktop.
Product for Windows 11
Your private AI assistant, on your PC. Albireo runs language models on the computer where it is installed: chat, documents, images, agent tools and an OpenAI-compatible interface, with nothing sent outside.

Product for Windows 11
Share your documents without exposing the personal data they contain. Pseudalis detects it and replaces each value with an encrypted tag that only the key of your computer can restore.
Open-source library
Transformer architectures for the Dlib C++ library. Import an open-weight model in GGUF format, run it, convert it into a trainable network, fine-tune it, distill it and deploy it.
Each figure comes from a public report or README of the repository and states the conditions of its measurement.
| Subject | Result | Conditions |
|---|---|---|
| GPU memory extended in software | A working set of 17,506 MiB trained on a card of 12,288 MiB | Model of 270 million parameters, batch of 8, budget of 8,192 MiB |
| Personal data detection | F1 of 0.946 in five languages; 2 of 1,223 values left readable on sixteen test files | Pseudalis 1.0.11: synthetic test set in five languages, exact spans, every category; Word, PowerPoint, PDF, CSV, source and image files written for the test |
| Generation on a laptop GPU | 56.8 tokens per second, against 23.6 on the processor | Qwen3-VL 4B in Q4_K_M, GeForce RTX 5070 Ti Laptop GPU, 24 processor threads |
| Lossless text compression | enwik9 in 117,169,624 bytes, 0.937 bits per byte | Decompressor of 7,063,888 bytes, tokenizer and network included |
| Training routed to the integrated GPU | 63.3 ms per step instead of 97.4 ms | Intel Core Ultra 9 275HX, automatic routing, final loss within 0.001 of the host's |
| Reading a prompt on a laptop | 216 tokens per second with Qwen3-VL 4B, more than 1,300 with Qwen3 0.6B | Intel Core Ultra 9 275HX, 24 threads, processor engine with its AVX-VNNI kernels, first pass of a prompt of 512 tokens |
Models, documents and keys stay on machines their owners control. The code can be read, audited, modified and self-hosted.
Small models, quantized weights and careful kernels let a laptop do work usually sent to a server.
Methods, measurements and limits are published with the code, in reports written to be checked and reproduced.
Our public repositories cover transformers and language models, tokenization, image colorization, generative adversarial networks and self-supervised vision.

Investors, distributors and sponsors: fund the roadmap, distribute the products under your own brand, or add them to your offering.