Our work and P16
P16 is the French programme, led by Inria and supported by France 2030, that develops sovereign, open and interoperable software libraries for AI. Our library and our products pursue the same objectives, and we intend to seek its support.
P16 in brief
Launched in September 2024 within the National Strategy for Artificial Intelligence, P16 takes its name from measure 16 of the AI acceleration strategy. Its mission is to strengthen the technological autonomy of France and Europe through software libraries that cover the whole cycle of data and models. Inria steers the programme through its AI Programme, with about 8 million euros over five years.
- Three priority actions: the interoperability of data, their preparation, and the training and execution of models.
- Libraries that are developed, widely distributed and maintained as digital commons.
- A path of transfer to industry, including for research on trusted, frugal and embedded AI.
- Software components selected on their quality, their maturity, their genericity and their interest for digital sovereignty.
Convergence between our work and the objectives of P16
Each line links an objective stated by P16 to what we have already published.
| Objective of P16 | Publications of Cydral Technology |
|---|---|
| Sovereign and open libraries | Dlib-Transformer-extensions, written in France in standard C++14 and published under the Boost Software License 1.0, with no Python environment in the chain. |
| Interoperability | Models read from the GGUF format, a local interface that follows the OpenAI interface and is described with OpenAPI, documents read and written in their office formats. |
| Data preparation | A byte-pair encoding tokenizer, corpus and dataset tools, and Pseudalis, which removes personal data from documents before they feed a model or a prompt. |
| Training and execution of models | Fine-tuning with LoRA and DoRA, distillation, pruning, and engines for processors, for the NPU and GPU of Intel and for NVIDIA graphics processors. |
| Frugal AI | Quantized weights, integer kernels and software-extended GPU memory: a multimodal model of 4 billion parameters answers on a laptop with 8 GB of memory. |
| Trusted AI | Local execution, signed updates, reversible pseudonymization under articles 25 and 32 of the GDPR, and reports that state the limits of each method. |
| Embedded AI | Operations routed to the NPU and to the integrated GPU of the processor, and models compiled into a single program. |
| Transfer to companies | Two installed products, Albireo and Pseudalis, free in their Community editions and open to distribution under another brand. |
Against its selection criteria
Software quality
Tests, runnable examples and a research report for each method, kept with the code.
Maturity
Released versions with their installers, a signed update channel and a model catalog checked by hash.
Genericity
One library covers the life of a model, from the import of a published model to its distillation, for text and for images.
Sovereignty
Code, models and keys stay on machines their owners control, and every component can be read, audited and hosted.
The support we seek
Cydral Technology is not a partner of P16 today and receives no public funding. We intend to present the library to the programme and to find there engineering support, visibility and, if possible, funding. This would keep the library maintained over time and bring it to more European companies.
Any other support is welcome meanwhile: sponsoring a feature, funding the roadmap or distributing the products.
Sources
- Inria: what is P16, the project entrusted to Inria within France 2030 (in French)
- The P16 website
- French National Research Agency, project P16 (23-RDIA-0001)
P16 and France 2030 are named here to describe public programmes. Their mention implies no endorsement of Cydral Technology by Inria or by the French State.