Our Large Language Model as a Service (LLMaaS) offering gives you access to cutting-edge language models, inferred using SecNumCloud-qualified infrastructure, HDS-certified for healthcare data hosting, and therefore sovereign, calculated in France. Benefit from high performance and optimal security for your AI applications. Your data remains strictly confidential, and is neither exploited nor stored after processing.
Simple, transparent pricing
1.8
€
per million input tokens
8
€
per million tokens issued
8
€
per million reasoning tokens
4
€
per million reranking tokens
0,01 €
per minute of transcribed audio *
Calculated on an infrastructure based in France, SecNumcloud qualified and HDS certified.
Note on the "Reasoning" price:
This price applies specifically to models classified as "reasoners" or "hybrids" (models with the "Reasoning" capability activated) when reasoning is active and only on tokens linked to this activity.
* any minute started is counted
Chat & Reasoning
Our large models offer state-of-the-art performance for the most demanding tasks. They are particularly well-suited to applications requiring a deep understanding of language, complex reasoning or the processing of long documents.
80 tokens/second
qwen3.6:27b
Generalist reference model with a native context of 1M tokens. Excels at reasoning, following instructions and multilingualism.
Significant improvements in following instructions, reasoning, reading comprehension, mathematics, coding and tool use. Its context of 1M tokens enables the analysis of entire documents without truncation.
94 tokens/second
gpt-oss:120b
OpenAI's state-of-the-art open-weight model with configurable reasoning and transparent chain of thought.
Mixture-of-Experts model with 120 billion parameters offering configurable reasoning and full access to the chain of thought. Ideal for scenarios requiring a permissive licence (Apache 2.0).
41 tokens/second
gpt-oss:20b
Compact version of the OpenAI model, optimised for rapid inference with good reasoning capabilities.
Mixture-of-Experts model with 21 billion parameters and 3.6 billion active parameters. Configurable reasoning and full agent capabilities.
10 tokens/second
llama3.3:70b
Multilingual Meta model, excellent in natural dialogue and nuanced understanding in 8 languages.
Supports English, French, German, Spanish, Italian, Portuguese, Hindi and Thai. Its 132k tokens window enables analysis of complex documents and long conversations.
23 tokens/second
gemma3:27b
Google multimodal model with integrated vision and support for 140+ languages. Context of 120K tokens.
Includes native multimodal capabilities (text + image) and excels in over 140 languages. Ideal for analysing large documents and document research.
72 tokens/second
nemotron-3-super:120b
NVIDIA model optimised for collaborative agents, long reasoning and high-volume workloads. 1M tokens context.
Ideal for agentic workflows, long-context reasoning, high-volume automation (support tickets, mass analyses), the use of tools and RAG.
160 tokens/second
nemotron3-nano:30b
Ultra-fast NVIDIA model (160 t/s) with reasoning and function calling. Context of 1M tokens.
Excels at function calling, structured reasoning and analysing long contexts. Rare combination of high speed and very long context.
130 tokens/second
nemotron-cascade:30b
NVIDIA model specialising in mathematics (IMO 2025 gold medal) and problem decomposition. Context 1M tokens.
Excels at structured reasoning, solving complex mathematical problems and analysing long contexts.
88 tokens/second
glm-4.7-flash:30b
Fast model with an excellent performance/latency balance for reasoning and analysis.
Offers fast inference (88 t/s) with a context of 120k tokens. Particularly suited to conversational assistants requiring low latency.
21 tokens/second
cogito:32b
Advanced analytical reasoning model, designed for complex problem decomposition and logic verification.
Excels in multi-factor analysis, formal demonstration and hallucination minimisation thanks to built-in logic verification mechanisms.
22 tokens/second
elm tree 3:32b
The first fully open reasoning model on this scale. Total transparency (data, code, weight).
Competes with the best proprietary models on complex benchmarks (MATH, HumanEval+). Able to expose its thought process. Preferred choice for transparency and auditability.
35 tokens/second
Elm 3:7b
Completely open and efficient model, excellent in mathematics and programming with total transparency.
Optimised for efficiency (2.5x less resources than Llama 3.1 8B). Ideal for tasks requiring complete reproducibility and auditability.
56 tokens/second
qwen3-2507:235b
The most powerful model in the catalogue (235B parameters, 22B active). Excels in maths, coding and logical reasoning.
Ultra-sparse Mixture-of-Experts architecture combining the power of a very large model with the efficiency of a smaller model.
28 tokens/second
mistral-small3.2:24b
Mistral model with improved instruction tracking, robust function calling and vision capabilities. Integrated problem content detection.
Excellent instruction tracking, fewer repetitions, reliable function calling. Supports vision (image analysis) and native security filters.
100 tokens/second
mistral-small4:119b
High-performance Mistral model (119B) with vision, integrated security and context of 262K tokens. Fast (100 t/s).
Large version of the Mistral Small family. Combines power, speed and reliability with an extended context. Native security filters.
28 tokens/second
ministral-3:14b
The most powerful of the Ministral family, with advanced reasoning and coding. Context of 250K tokens.
Excels at complex reasoning and coding while remaining efficient.
40 tokens/second
ministral-3:8b
Intermediate Ministral model with an excellent performance/speed ratio. Context of 250K tokens.
Capable of complex reasoning while remaining fast. Ideal for assistants requiring responsiveness and quality.
22 tokens/second
ministral-3:3b
Compact Mistral model, high performance despite its small size. Context of 250K tokens.
Surprising performance for conversational tasks and simple reasoning despite only 3B parameters.
32 tokens/second
qwen3.5:9b
Qwen3.5 intermediate model with solid reasoning and context extended to 250K tokens.
Good balance between generation quality and inference speed.
37 tokens/second
qwen3.5:4b
Compact Qwen3.5 model with a good performance/efficiency ratio and a 250K token context.
Good candidate for assistants and light reasoning tasks.
16 tokens/second
qwen3.5:0.8b
Ultra-light model with an exceptional background of 250K tokens - remarkable for a model of this size.
Ideal for fast conversational tasks requiring a very long history or analysis of large documents with a small footprint.
46 tokens/second
qwen3:0.6b
Ultra-fast micro-model for simple tasks and routing. 40K context tokens.
Ideal as the first level of processing in complex workflows or for rapid classification tasks.
55 tokens/second
qwen3-2507-think:4b
Compact model optimised for deep reasoning (logic, maths, science, code). Context of 250K tokens.
Thinking" version with enhanced reasoning capability. Combines compactness, speed and advanced reasoning.
19 tokens/second
qwen3-omni:30b
Native omnimodal model - includes text, image, video and audio simultaneously.
Supports multimodal input (text, image, audio, video) with advanced reasoning capabilities. Note - audio output via API is not yet enabled.
Programming & Agents
Our programming and agent models are specially optimised for agentic software engineering, large-scale code generation and development workflow automation.
121 tokens/second
qwen3.6:35b
Leader in agentic software engineering (SWE-bench 73.4%). Context of 1M tokens, integrated vision and tool calling.
Includes entire code repositories thanks to its 1M token context. Supports multi-step reasoning and vision (screenshots, diagrams). Optimised for IDEs and CI/CD pipelines.
97 tokens/second
qwen-coder-next:80b
State-of-the-art model for complex code and reasoning. Context of 250K tokens.
Excels at large-scale code generation and analysis. Designed for advanced software engineering tasks.
67 tokens/second
qwen3-next:80b
Versatile 80B model optimised for large contexts, function calling and structured reasoning.
Context of 250K tokens with support for function calling and guided decoding.
33 tokens/second
devstral-small-2:24b
State-of-the-art agentic model for software engineering. Performance close to >100B models for code. Integrated vision.
Optimised for codebase exploration, multi-file editing and the use of tools. Native vision support. Context of 200K tokens.
23 tokens/second
rnj-1:8b
Specialised STEM model - excels in code (83.5% HumanEval+), maths and science.
Dense model trained on 8.4T tokens. Often outperforms much larger models on code and mathematical reasoning tasks.
40 tokens/second
functiongemma:270m
Micro-model specialising in function call detection. Ideal as a router in an agentic architecture.
Ultra-compact, optimised for identifying and formatting function calls quickly.
Vision & Multimodal
Our Vision & Multimodal models can analyse images, videos and visual documents. They excel in OCR, object detection, structured extraction and spatio-temporal reasoning.
24 tokens/second
qwen3-vl:235b
The most powerful multimodal model in the catalogue. Advanced visual understanding and exceptional reasoning.
Excels in complex document analysis, multilingual OCR, 3D spatial reasoning and video understanding.
17 tokens/second
qwen3-vl:32b
High-performance variant for the most demanding vision tasks. Context 250K tokens.
Fine analysis of high-resolution images, understanding of dynamic scenes and text-timestamp alignment for video.
39 tokens/second
qwen3-vl:30b
High-performance multimodal model for OCR, object detection, video analysis and spatio-temporal reasoning.
Incorporates innovations in image and video analysis. Excels in complex OCR, graphics and structured extraction (JSON).
39 tokens/second
qwen3-vl:8b
Intermediate vision model - good compromise between performance and footprint. Context 250K tokens.
Capable of analysing complex documents, graphics and videos with a high degree of accuracy.
57 tokens/second
qwen3-vl:4b
Compact, fast vision model for document analysis and video comprehension.
Excellent compromise between performance and footprint. Supports structured extraction and visual reasoning.
64 tokens/second
qwen3-vl:2b
Ultra-compact vision model for rapid OCR, object detection and embedded applications.
Despite its small size, it offers amazing image and video analysis. Ideal for mobile or embedded applications.
59 tokens/second
gemma4:31b
Google's dense multimodal model, ranked 3rd in the world on Arena AI. Advanced vision, reasoning and coding. Context 250K tokens.
Google's most powerful open-source model. Native function calling, advanced visual understanding (OCR, graphics, documents, UI). Multilingual (35+ languages).
125 tokens/second
gemma4:e2b
Ultra-fast variant (125 rpm) of Gemma 4 with vision. Excellent energy efficiency.
Offers an exceptional performance to footprint ratio. 128K tokens with full vision capabilities.
85 tokens/second
gemma4:e4b
Gemma 4 variant with a better quality/speed compromise than the E2B version. Integrated vision.
Better fidelity than the E2B version, but still high speed. Context 128K tokens.
49 tokens/second
granite3.2-vision:2b
IBM Granite compact vision model for rapid OCR and data extraction from scanned documents.
Lightweight yet powerful for low-latency OCR and image analysis.
84 tokens/second
deepseek-ocr
Specialised OCR model for high-precision text extraction with preserved formatting (tables, formulas).
Optimised for converting documents into structured Markdown. Excels with complex tables and formulas.
Embedding
Our embedding models transform text into vector representations for semantic search, clustering and RAG (Retrieval-Augmented Generation) pipelines.
171 tokens/second
bge-m3:567m
State-of-the-art multilingual embedding (100+ languages). Supports dense, sparse and multi-vector searches.
Context of 8192 tokens with three complementary search methods.
qwen3-embedding:4b
High-performance embedding with deep semantic understanding and extended context (40K tokens).
Ideal for processing large documents in RAG pipelines.
qwen3-embedding:8b
High-capacity embedding with the best semantic understanding of the Qwen3 family. Extended context (40K tokens).
The most powerful version of the Qwen3 embedding family. Ideal for tasks requiring contextual understanding.
qwen3-embedding:0.6b
Ultra-light and fast embedding for low-latency semantic search.
Excellent compromise between semantic performance and speed of execution.
196.3 tokens/second
granite-embedding:278m
Ultra-compact IBM embedding for semantic search with minimal latency.
The fastest embedding model in the catalogue. Ideal for clustering and high-frequency searching.
175 tokens/second
embeddinggemma:300m
Multilingual Google embedding (100+ languages), optimised for search and semantic retrieval.
Produces vector representations of text for classification, clustering and similarity search.
Reranking
Our reranking models reorder search results by relevance to refine the quality of RAG pipelines. Compatible with the Cohere API.
nvidia/llama-nemotron-rerank-vl-1b-v2
Cohere API-compatible reranking model (/v1/rerank and /v2/rerank). Orders documents by relevance to a query.
Cohere v1/v2 SDK compatible. The relevance score is a raw logit (relative order is guaranteed). Ideal as a complement to the RAG stack (embedding + retrieval + rerank).
qwen3-reranker:4b
Powerful reranking model with a high level of contextual understanding.
Excellent rescheduling quality thanks to its 4B parameters. Ideal for demanding RAG pipelines.
qwen3-reranker:0.6b
Compact and efficient reranking model for rapid rescheduling.
Lightweight version for use cases requiring low reranking latency.
bge-reranker-large
High-performance multilingual reranking model from the BGE family.
Complementary to the BGE-M3 embedding model for complete RAG pipelines.
Security
Our security models specialise in detecting problematic content, preventing jailbreaks and ensuring regulatory compliance (RGPD, HDS). They can be used as pre-filters or post-filters in your workflows.
45 tokens/second
granite3-guardian:8b
Security model specialising in detecting problematic content, jailbreaking and regulatory compliance.
Designed to filter sensitive content and ensure RGPD/HDS compliance. Can be used as a pre-filter or post-filter in your workflows.
60 tokens/second
granite3-guardian:2b
Compact version of the Granite Guardian security model for low-latency filtering.
Same filtering capabilities as version 8B, but with a smaller footprint. Ideal for high-frequency workflows.
Translation
Our translation models offer high fidelity in 55 languages, respecting the grammar, cultural nuances and technical specificities of the documents.
17 tokens/second
translategemma:27b
High-performance translation for 55 languages. Superior quality for complex and technical content.
Captures literary and cultural nuances with exceptional fidelity.
27 tokens/second
translategemma:12b
High-fidelity translation for 55 languages with 128K tokens.
Respects grammar and cultural nuances. Ideal for long documents.
31 tokens/second
translategemma:4b
Fast, efficient translation for 55 languages. Ideal for real-time localisation.
Compact version with an excellent speed/quality ratio. Context of 128K tokens.
Audio & Image
Our Audio & Image models enable real-time voice transcription (ASR streaming) and image generation from text descriptions, compatible with the OpenAI API.
voxtral
Real-time audio transcription via WebSocket. Low-latency streaming speech recognition.
Operates in Realtime mode via the /v1/realtime endpoint (WebSocket). Transcribes streaming audio.
z-image:16b
Image generation from text prompts, OpenAI API compatible /v1/images/generations.
Supports image size and number of images. Compatible with the OpenAI ecosystem.
Model comparison
This comparison table will help you choose the model best suited to your needs, based on various criteria such as context size, performance and specific use cases.
| Model | Publisher | Parameters | Context (k tokens) | Vision | Agent | Reasoning | Security | Quick * | Energy efficiency * |
|---|---|---|---|---|---|---|---|---|---|
| Chat & Reasoning | |||||||||
| qwen3.6:27b | Qwen Team | 27B | 1000000 | ||||||
| gpt-oss:120b | OpenAI | 120B | 120000 | ||||||
| gpt-oss:20b | OpenAI | 20B | 120000 | ||||||
| llama3.3:70b | Meta | 70B | 132000 | ||||||
| gemma3:27b | 27B | 120000 | |||||||
| nemotron-3-super:120b | NVIDIA | 120B | 1000000 | ||||||
| nemotron3-nano:30b | NVIDIA | 30B | 1000000 | ||||||
| nemotron-cascade:30b | NVIDIA | 30B | 1000000 | ||||||
| glm-4.7-flash:30b | Zhipu AI | 30B | 120000 | ||||||
| cogito:32b | Deep Cogito | 32B | 32000 | ||||||
| elm tree 3:32b | AllenAI | 32B | 65536 | ||||||
| Elm 3:7b | AllenAI | 7B | 65536 | ||||||
| qwen3-2507:235b | Qwen Team | 235B | 200000 | ||||||
| mistral-small3.2:24b | Mistral AI | 24B | 128000 | ||||||
| mistral-small4:119b | Mistral AI | 119B | 262144 | ||||||
| ministral-3:14b | Mistral AI | 14B | 250000 | ||||||
| ministral-3:8b | Mistral AI | 8B | 250000 | ||||||
| ministral-3:3b | Mistral AI | 3B | 250000 | ||||||
| qwen3.5:9b | Qwen Team | 9B | 250000 | ||||||
| qwen3.5:4b | Qwen Team | 4B | 250000 | ||||||
| qwen3.5:0.8b | Qwen Team | 0.8B | 250000 | ||||||
| qwen3:0.6b | Qwen Team | 0.6B | 40000 | ||||||
| qwen3-2507-think:4b | Qwen Team | 4B | 250000 | ||||||
| qwen3-omni:30b | Qwen Team | 30B | 32768 | ||||||
| Programming & Agents | |||||||||
| qwen3.6:35b | Qwen Team | 35B | 1000000 | ||||||
| qwen-coder-next:80b | Qwen Team | 80B | 250000 | ||||||
| qwen3-next:80b | Qwen Team | 80B | 250000 | ||||||
| devstral-small-2:24b | Mistral AI & All Hands AI | 24B | 200000 | ||||||
| rnj-1:8b | Essential AI | 8B | 32000 | ||||||
| functiongemma:270m | 270M | 32768 | |||||||
| Vision & Multimodal | |||||||||
| qwen3-vl:235b | Qwen Team | 235B | 200000 | ||||||
| qwen3-vl:32b | Qwen Team | 32B | 250000 | ||||||
| qwen3-vl:30b | Qwen Team | 30B | 250000 | ||||||
| qwen3-vl:8b | Qwen Team | 8B | 250000 | ||||||
| qwen3-vl:4b | Qwen Team | 4B | 250000 | ||||||
| qwen3-vl:2b | Qwen Team | 2B | 250000 | ||||||
| gemma4:31b | 31B | 250000 | |||||||
| gemma4:e2b | 31B (E2B) | 128000 | |||||||
| gemma4:e4b | 31B (E4B) | 128000 | |||||||
| granite3.2-vision:2b | IBM | 2B | 16384 | ||||||
| deepseek-ocr | DeepSeek AI | 3B | 8192 | ||||||
| Embedding | |||||||||
| bge-m3:567m | BAAI | 567M | 8192 | ||||||
| qwen3-embedding:4b | Qwen Team | 4B | 40000 | ||||||
| qwen3-embedding:8b | Qwen Team | 8B | 40000 | ||||||
| qwen3-embedding:0.6b | Qwen Team | 0.6B | 32768 | ||||||
| granite-embedding:278m | IBM | 278M | 512 | ||||||
| embeddinggemma:300m | 300M | 2048 | |||||||
| Reranking | |||||||||
| nvidia/llama-nemotron-rerank-vl-1b-v2 | NVIDIA | 1B | 4096 | N.C. | |||||
| qwen3-reranker:4b | Qwen Team | 4B | 4096 | N.C. | |||||
| qwen3-reranker:0.6b | Qwen Team | 0.6B | 4096 | N.C. | |||||
| bge-reranker-large | BAAI | 335M | 512 | N.C. | |||||
| Security | |||||||||
| granite3-guardian:8b | IBM | 8B | 8192 | ||||||
| granite3-guardian:2b | IBM | 2B | 8192 | ||||||
| Translation | |||||||||
| translategemma:27b | 27B | 120000 | |||||||
| translategemma:12b | 12B | 128000 | |||||||
| translategemma:4b | 4B | 128000 | |||||||
| Audio & Image | |||||||||
| voxtral | Mistral AI | 4B | 32768 | N.C. | |||||
| z-image:16b | Community | 16B | N.C. | ||||||
Legend and explanation
Functionality or capacity supported by the model
Functionality or capability not supported by the model
* Energy efficiency
Indicates particularly low energy consumption (< 2.0 kWh/Mtoken)
* Quick
Model capable of generating more than 50 tokens per second
Note on performance measures
The speed values (tokens/s) represent performance targets in real-life conditions. Energy consumption (kWh/Mtoken) is calculated by dividing the estimated power of the inference server (in Watts) by the measured speed of the model (in tokens/second), then converted into kilowatt-hours per million tokens (division by 3.6). This method offers a practical comparison of the energy efficiency of different models, to be used as a relative indicator rather than an absolute measure of power consumption.
Recommended use cases
Here are some common use cases and the most suitable models for each. These recommendations are based on the specific performance and capabilities of each model.
Multilingual dialogue
Chatbots and assistants able to communicate in several languages with automatic detection and context maintenance
Recommended models
- nemotron-3-super:120b
- qwen3.6:27b
- nemotron3-nano:30b
- gpt-oss:120b
Analysis of long documents
Processing of large documents (>100 pages) with extraction of key information, summaries and answers to questions
Recommended models
- nemotron-3-super:120b
- qwen3.6:27b
- qwen3-2507:235b
Programming and development
Code generation, optimisation and debugging in multiple languages, refactoring and test creation
Recommended models
- qwen3.6:35b
- qwen-coder-next:80b
- devstral-small-2:24b
- nemotron-3-super:120b
Visual analysis
Image and visual document processing, OCR, interpretation of graphs and tables
Recommended models
- qwen3-vl:235b
- gemma4:31b
- deepseek-ocr
- qwen3-vl:30b
Safety and compliance
Sensitive content filtering, jailbreak detection, RGPD/HDS compliance
Recommended models
- granite3-guardian:8b
- granite3-guardian:2b
- mistral-small4:119b
Light deployments
Applications requiring a minimal footprint, low latency and low power consumption
Recommended models
- qwen3.5:0.8b
- qwen3-vl:2b
- ministral-3:3b
RAG (Retrieval-Augmented Generation)
Complete semantic search, reordering and retrieval-enhanced generation pipelines
Recommended models
- bge-m3:567m
- nvidia/llama-nemotron-rerank-vl-1b-v2
- qwen3.6:27b