The best open source AI on demand in a sovereign cloud

Designed for complex tasks that require a broad context and greater precision in logical reasoning.

An optimised architecture for faster inference and reduced power consumption, freeing up significant computational resources.

Trained on millions of agents and tasks of increasing complexity to ensure robust adaptability in the real world.

Ideal for multilingual services, government agencies and R&D teams looking for a reliable, adaptable model

Data and methods documented for unprecedented transparency

Compliant with the AI Act and respectful of privacy and intellectual property

A 70B version with performance on a par with current market leaders

The perfect balance between responsiveness and power, designed to excel at logical reasoning, in-depth document analysis and the generation of reliable code.

Leverages cutting-edge architecture to provide a nuanced understanding of broad contexts and complex instructions.

Ideal for advanced chatbots and enterprise workflows that require high flexibility without compromising on processing speed.

Native multimodal: converts text, images or mockups into fully functional code.

Designed for large-scale development: includes an extended context window of up to 256k tokens to manage complex projects

Optimised for vibe coding: a fast, fluid and creative experience designed for developers and product designers

Compatible with agent-based workflows: automates analysis, code generation, and end-to-end execution

Optimised for fast and cost-effective deployment, ideal for chatbots, document analysis and specialised tasks.

Offers performance comparable to Mistral Small 3.2 24B with minimal resources.

Capable of analysing images and providing information based on visual content, in addition to text.

A state-of-the-art multilingual model capable of processing short queries, paragraphs and long documents of up to 8192 tokens simultaneously

Combines lexical (keywords) and semantic (meaning) analysis for unrivalled ranking accuracy on complex data sets

The ideal solution for enterprise search engines and RAG applications that require a deep understanding of context

Ultra-lightweight architecture (0.6 billion parameters) designed for ultra-low-latency inference and minimal power consumption

Maintains high relevance accuracy, even with a context window expanded up to 32768 tokens

Ideal for real-time data feeds, autonomous agents and large-scale deployments

The most powerful open-source embedding model on the market

The benchmark for semantic search and augmented search (RAG) tasks

Ideal for advanced use of embedding vectors in a variety of use cases

Outstanding performance, whatever language the text is in (100+ languages)

This model is the result of community work based on a model published by Microsoft.

Excellent value for money, perfect for prototyping and simple tasks with limited resources

Great performance for relatively simple tasks, whatever language the text is in

Extreme speed for indexing huge databases or real-time processing

High energy efficiency to reduce environmental impact

Model trained on over 1 million hours of data

Transcription errors reduced by up to 20% compared with Whisper V2

Better handling of accents, background noise and complex speech (e.g., calls or videoconferences)

Enhanced multilingual support and translation of transcriptions into languages other than English

The best combination of quality and speed in generative AI image creation

Fast generation of photo-realistic images in 1, 2, 4 or 8 steps based on prompts

Operates by distillation, which increases energy efficiency and ensures excellent quality

Optimised for English, with limited knowledge of other languages (FR, DE, ES, IT, etc.)

Create photos in multiple styles from one or more profile photos

Powerful and flexible: recontextualisation, colourisation, age and gender change, mix of identities, etc.