Choose among ready-to-be-served AI models
What makes inference fast? Among many things, model optimization. That's why Scaleway is providing an evolutionary Model Library, offering curated and quantized models, including LLMs and embeddings.
Managed Inference
Serve Generative AI models and answer prompts from European end-consumers securely. A drop-in replacement for all apps using OpenAI APIs.
Enjoy unlimited tokens at predictable price
No matter how big your usage is, you pay the same —predictable— price for unlimited tokens. This price depends on the dedicated infrastructure that serves your model, which is billed per hour.
Run on a fully secured European Cloud
Maintain complete data control: your prompts and responses are not stored and cannot be accessed by Scaleway or any third parties, keeping your data exclusively yours and within Europe.
Open-weights language and embedding models
Pixtral-12b-2409
Vision language model able to analyze your images and offer insights without compromising on instruction following. Another fantastic model made by Mistral AI and distributed under the Apache 2.0 license.
Llama-3.1-8b-instruct
Llama 3.1 by Meta is the latest iteration of the open-access Llama family, for efficient deployment and development on smaller GPUs. Llama models are tailored for dynamic dialogues and creative text generation. Engineered with the latest in efficiency and scalability, it excels in complex reasoning and coding tasks. Its advanced Grouped-Query Attention mechanism ensures unparalleled processing prowess, making it the ultimate tool for chat applications and beyond.
Llama-3.1-70b-instruct
Llama-3.1-70b-Instruct by Meta is the most powerful model in the Llama family, boasting 70 billion parameters for unmatched text generation. Ideal for advanced AI applications, it excels in multilingual dialogues, complex reasoning, and intricate coding tasks. Its Grouped-Query Attention mechanism ensures efficient processing, making it the top choice for high-end deployment.
Mistral-Nemo-Instruct
A 12B model with a very large context window of up to 128k tokens, particularly useful for RAG applications. Built by Mistral in collaboration with NVIDIA, this model is distributed under the Apache 2.0 license.
Mixtral-8x7B-Instruct-v0.1
Trained on Scaleway's Nabuchodonosor 2023, Mixtral-8x7B is a state-of-the-art, pretrained generative model known as a Sparse Mixture of Experts. It has been benchmarked to surpass the performance of the Llama 2 70B model across a variety of tests.
BGE-Multilingual-Gemma2
An advanced embedding model that translates data into vectors, capturing complex relationships for enhanced information processing. Perfect to set up your Retrieval-Augmented Generation (RAG) system.
Try the 1st European Inference product for generative AI models on the market
Predictable pricing
Pick among on-the-shelf optimized models, and get a dedicated inference endpoint right away.
You are charged for usage of the GPU type you choose.
| Model | Quantization | GPU | Price | Approx. per month |
| Llama3.1-8b-instruct | BF16, FP8 | L4-1-24G | €0.93/hour | ~€679/month |
| Llama3.1-70b-instruct | FP8 | H100-1-80G | €3.40/hour | ~€2482/month |
| Llama3.1-Nemotron-70b-instruct | FP8 | H100-1-80G | €3.40/hour | ~€2482/month |
| Mistral-7b-instruct-v0.3 | BF16 | L4-1-24G | €0.93/hour | ~€679/month |
| Pixtral-12b-2409 | BF16 | H100-1-80G | €3.40/hour | ~€2482/month |
| Mistral-nemo-instruct-2407 | FP8 | H100-1-80G | €3.40/hour | ~€2482/month |
| Mixtral-8x7b-instruct-v0.1 | FP8 | H100-1-80G | €3.40/hour | ~€2482/month |
| BGE-Multilingual-Gemma2 | FP32 | L4-1-24G | €0.93/hour | ~€679/month |
More models and conditions available on this page.
Benefit from a secured European Cloud ecosystem
Virtual Private Cloud
Your AI endpoints are accessible through low-latency and secure connection to your resources hosted at Scaleway, thanks to a resilient regional Private Network.
Access Management
We make generative AI endpoints compatible with Scaleway's Identity and Access Management, so that your deployments are compliant with your enterprise architecture requirements.
Cockpit
Identify bottlenecks on your deployments, view inference requests in real time and even report your energy consumption with a fully managed observability solution.
Frequently asked questions
How can I start using this service?
You'll find here a comprehensive guide on getting started, including details on deployment, security, and billing.
If you need support, don't hesitate to reach out to us through the dedicated slack community #inference-beta
What are Scaleway's security protocols for AI services?
Scaleway’s AI services implement robust security measures to ensure customer data privacy and integrity. Our measures and policies are published on our documentation.
Can I use the OpenAI libraries and APIs?
Scaleway lets you seamlessly transition applications already utilizing OpenAI. You can use any of the OpenAI official libraries, for example the OpenAI Python client library, to interact with your Scaleway Managed Inference deployments. Find here the APIs and parameters supported.
What are the advantages over mutualized LLM API services?
- Complete isolation of computing and networking resources to ensure maximum control for sensitive applications.
- Consistent and predictable performance, unaffected by the activity of other users.
- No strict rate limits—usage is only constrained by the maximum load your deployment can handle.
- Access to a wider range of models.
- More cost-effective with high utilization.
Do you have pay-per-token hosted models?
Managed Inference deploys AI models and creates dedicated endpoints on a secure production infrastructure.
Alternatively, Scaleway has a selection of hosted models in its datacenters, priced per million tokens consumed, available via API. Find all details on the Generative APIs page.
I've got a request, where can I share it?
Tell us the good and the bad about your experience here. Thank you for your time!
Get started with tutorials
Retrieval-Augmented Generation (RAG)Learn how to implement RAG using LangChain and Managed Inference
Processing images with a vision modelGetting structured outputs with Pixtral vision model
Get started with agentic AIUse function calling on open-weight Llama 3.1