Scaleway

What is Scaleway?

Scaleway is a French cloud provider headquartered in Paris and part of the Iliad Group. Its broader cloud platform includes compute, object storage, managed Kubernetes, networking and GPU infrastructure, so its AI services sit within a complete infrastructure portfolio rather than a standalone model marketplace.

Its AI offering has two layers. Generative APIs provide serverless, pay-per-token access without infrastructure management, while Managed Inference supports dedicated deployments, guaranteed throughput, private network isolation, custom quantization and imported models. Eden AI exposes the Generative APIs layer. Scaleway hosts open-weight models from external developers rather than training its own foundation models.

Scaleway at a glance

Scaleway main AI capabilities

• Chat and text generation: Produces conversational answers, drafts, classifications and general-purpose completions.
• Reasoning: Handles multi-step analysis using models such as GLM 5.2 and Qwen3.5.
• Code generation: Generates and edits code with Devstral 2 and Qwen3 Coder.
• Vision-language understanding: Interprets images alongside text using Pixtral 12B.
• Audio and speech understanding: Transcribes and summarizes audio using Voxtral Small 24B.
• Computer-use and GUI agents: Navigates web, desktop and mobile interfaces using Holo2.
• Text embeddings: Converts text into vectors for search, clustering and retrieval.
• Structured JSON and function calling: Returns schema-constrained data and invokes application tools.
• Batch processing: Runs non-real-time workloads at discounted rates through Scaleway’s batch API.

When should you choose Scaleway?

Choose Scaleway when your sovereignty requirement goes beyond locating servers inside the EU. A French company operating under French and EU law creates a different jurisdictional position from a US-controlled provider with European data centers, particularly when procurement teams are assessing US CLOUD Act exposure.

It is also relevant for healthcare and public-sector workloads where certification is a contractual requirement. Scaleway holds ISO 27001 and France’s HDS health data hosting certification, while its SecNumCloud qualification is under review by ANSSI and SGDSN.

Existing Scaleway customers can place inference beside compute, storage and Kubernetes workloads under the same cloud account, billing relationship and private network. They can begin with serverless Generative APIs and later move to Managed Inference for dedicated capacity or network isolation.

Scaleway is especially strong when you need Mistral models such as Mistral Medium 3.5, Mistral Small 3.2, Devstral 2, Pixtral or Voxtral on French-operated infrastructure.

Teams requiring GPT, Claude, Gemini or regional endpoints outside Europe should choose another provider.

Scaleway pros and cons

Scaleway models, features and capabilities on Eden AI

Relevant selected features for Scaleway

• Chat: Builds conversational assistants with multi-turn context and role-based messages.
• Text Generation: Produces drafts, classifications, transformations and general-purpose completions.
• Code Generation: Generates, explains, reviews and modifies source code.
• Multimodal Chat: Combines images, screenshots or audio with text instructions.
• Embeddings: Creates vectors for semantic search, retrieval and clustering.
• Summarization: Condenses documents, conversations or audio into structured summaries.
• Question Answering: Generates answers from prompts, supplied context or retrieved content.

Available Scaleway models on Eden AI

Eden AI exposes 17 Scaleway models covering general-purpose generation, coding, multimodal understanding, audio, computer use and embeddings. Context windows range from 22K to 256K tokens, while input pricing ranges from $0.12 to $2.08 per million tokens. Every listed model runs in the EU region. Model identifiers follow the format scaleway/<model-name>.

Frontier and general purpose

This group suits demanding general-purpose workloads. qwen3.5-397b-a17b is the practical default when you need broad capability without the highest input price.

Efficient and high volume

These models target production volume and cost-sensitive applications. mistral-small-3.2-24b-instruct is the default for balanced European-language performance, context and output cost.

Coding

This group supports code generation, repository work and developer agents. devstral-2-123b-instruct-2512 is the default for more demanding software engineering tasks.

Specialized

These models cover inputs and actions beyond text. pixtral-12b-2409 is the broadest default, while Voxtral and Holo2 should be selected for their specific modalities.

Embeddings

These models support semantic retrieval, clustering and vector search. bge-multilingual-gemma2 is the default for multilingual applications, while Qwen3 suits general embedding workloads.

Two models you will not find on most providers

Holo2 and Voxtral give the Scaleway catalog capabilities beyond standard text generation. Holo2, developed by the French company H Company, localizes interface elements and navigates real web, desktop and mobile applications. Its availability through a standard chat API makes Scaleway one of the few providers where you can test a computer-use agent without adopting a separate interaction protocol. Its 22K context window is a meaningful constraint for long workflows or extensive visual histories.

Voxtral Small 24B, developed by Mistral, processes spoken content rather than merely generating text. It supports speech understanding, transcription and structured audio summarization, making it suitable for meeting analysis, call processing and voice-based document extraction.

Supported Scaleway capabilities

Supported AI categories

• ‍Text / Generative AI: Chat, reasoning, summarization, extraction, question answering and code generation.‍
• Multimodal: Pixtral interprets images and screenshots alongside natural-language prompts.‍
• Audio: Voxtral processes speech for transcription, understanding and structured summarization.‍
• Embeddings: Qwen3 Embedding and BGE Multilingual Gemma2 generate vectors for retrieval and similarity search.

Scaleway API output: what data can be extracted or generated?

Important note on Scaleway accuracy and reliability

Open-weight model quality varies more substantially between individual models than quality usually varies within one proprietary model family. You should benchmark each Scaleway model against your own prompts, languages and output constraints rather than treating the catalog as interchangeable.

Several model identifiers include version date suffixes, such as -2409, -2507 and -2512. Pin the exact identifier in production so an integration does not silently depend on a different model revision.

Holo2 and Voxtral are specialized by design. Holo2 should be evaluated on interface localization and action planning, while Voxtral should be tested on transcription and audio understanding, not compared directly with general chat models.

Scaleway inference is currently hosted in Paris. Applications requiring geographic redundancy should configure cross-provider failover through Eden AI rather than assume multi-region resilience within Scaleway.

What can you build with Scaleway?

Scaleway supports regulated assistants, multilingual retrieval systems, coding tools and computer-use agents on French-operated infrastructure. The main design choice is selecting the right model for each workload rather than routing every request to one general-purpose model.

Use case 1 : GDPR-native assistants for regulated industries

A regulated assistant can retrieve approved internal documents, answer employee or customer questions, summarize case files and return structured JSON for downstream systems. Healthcare teams can use the workflow for clinical administration or patient-support content where HDS-certified hosting is a procurement requirement, while financial and public-sector teams may value Scaleway’s French jurisdiction and lack of US CLOUD Act exposure.

For routine traffic, scaleway/mistral-small-3.2-24b-instruct offers a 128K context window at $0.17 per million input tokens and $0.40 per million output tokens. Harder analysis can route to scaleway/mistral-medium-3.5-128b, which provides 256K context. Scaleway states that it does not collect, read, reuse or analyse prompt and output content.

At 100 million input tokens, Mistral Small costs $17 before output charges. Mistral Medium’s output price is $8.65 per million tokens, five times its $1.73 input rate, so verbose responses require strict output limits.

Use case 2: European RAG and multilingual semantic search

A European RAG system converts policies, contracts, support material or technical documentation into vectors, stores them in a vector database, retrieves relevant passages for each query and sends those passages to a generation model for synthesis. scaleway/bge-multilingual-gemma2 is the stronger default for multilingual European corpora where queries and source documents may mix French, German, Spanish, Italian or English. scaleway/qwen3-embedding-8b fits general semantic retrieval and should be benchmarked when language coverage is narrower.

Both embedding models cost $0.12 per million input tokens. Indexing 100 million source tokens therefore costs $12 in model input fees, excluding storage and re-indexing. Retrieval results can be passed to scaleway/glm-5.2, scaleway/mistral-medium-3.5-128b or scaleway/qwen3.6-35b-a3b, each offering a 256K context window for synthesis.

Embedding dimensions are not specified in the available catalog data, so confirm vector size before choosing an index schema. Larger context also does not replace retrieval quality or document-level access controls.

Use case 3: Coding assistants and computer-use agents

A coding assistant can inspect selected repository files, generate patches, explain failures and propose tests. scaleway/devstral-2-123b-instruct-2512 is suited to demanding software-engineering tasks with a 200K context window, while scaleway/qwen3-coder-30b-a3b-instruct offers a lower-cost option at $0.23 per million input tokens and $0.92 per million output tokens.

The same application can route interface tasks to scaleway/holo2-30b-a3b. Holo2, developed by H Company, identifies interface elements and navigates real web, desktop and mobile environments, allowing a computer-use agent to operate through a standard chat API. A practical router can send repository reasoning to Devstral, routine code generation to Qwen3 Coder and screenshot-driven actions to Holo2.

Holo2’s 22K context window limits how much interface history, screenshot metadata and prior action state can be passed at each step. Developers should compress state, retain only relevant observations and add deterministic confirmation checks before consequential actions.

Scaleway use cases by industry

Why use Scaleway through Eden AI?

Key benefits of using Scaleway on Eden AI

• One integration for Scaleway and alternative providers: Connect once to Eden AI, then access Scaleway alongside other European and global providers without maintaining separate authentication, request formats or SDK implementations.
• Normalized responses across providers: Eden AI converts provider-specific outputs into a consistent schema, reducing the application changes required when you compare models, add fallback or replace a deprecated model.
• Model switching through configuration: Moving from scaleway/mistral-small-3.2-24b-instruct to another provider or model becomes a parameter change rather than a new integration project.
• Faster replacement of superseded models: Open-weight models are frequently replaced by newer releases within a few months. A unified API reduces the engineering cost of testing and adopting those replacements.
• Centralized routing, monitoring and billing: You can manage provider selection, fallback order, usage and costs from one layer instead of operating separate integrations and commercial accounts.

Compare Scaleway with other AI models

Teams selecting a European provider often need to establish whether their jurisdictional requirements create a measurable trade-off in answer quality, coding performance or instruction following. Eden AI lets you run the same evaluation prompts through Scaleway models and non-European alternatives using the same request structure.

For example, you can compare scaleway/mistral-medium-3.5-128b with a proprietary frontier model on a held-out set of legal summaries, support questions or extraction tasks. This produces task-specific evidence rather than a general assumption that European hosting necessarily means lower quality. The result may justify using Scaleway everywhere, routing only difficult requests elsewhere, or retaining a strict EU-only policy despite a measured quality difference.

Add fallback and routing for production reliability

Scaleway inference is currently hosted in Paris, so Scaleway alone does not provide geographic redundancy across multiple inference regions. Eden AI lets you configure a second provider as fallback without building and maintaining another API integration. If Scaleway times out or becomes unavailable, the same normalized request can be routed to the next approved provider.

The fallback chain can remain entirely within Europe. You can configure Scaleway first, followed by another provider that meets your regional and contractual requirements, so failover does not silently send prompts to infrastructure outside the jurisdiction selected by your security or legal team. This distinction matters for healthcare, public-sector and financial workloads, where an operational fallback must not override the original data-processing policy. Provider order, model selection and retry behavior should be treated as part of the compliance configuration.

Monitor usage and costs across providers

Eden AI provides per-model and per-project visibility into usage and spending across Scaleway and any other providers used by the application. This is more useful than reviewing isolated provider dashboards when different models handle routine requests, difficult prompts, embeddings and fallback traffic.

Teams can track how model routing affects average request cost, identify projects producing unexpectedly verbose outputs and compare spend before and after a model change. Billing is consolidated into one invoice rather than separate invoices and payment relationships for every provider. This reduces reconciliation work while preserving enough detail to allocate usage internally by project, model or workload.

Best Scaleway alternatives and comparisons on Eden AI

Scaleway vs Nebius

Scaleway and Nebius both provide European infrastructure for open-weight model inference, but they solve different procurement problems. Scaleway’s stronger argument is legal and certification driven: it is a French company operating under European jurisdiction, outside US CLOUD Act reach, with ISO 27001 and HDS certification. Its SecNumCloud qualification is also under review by ANSSI, and its catalog provides particularly deep coverage of Mistral models.

Nebius is stronger when technical flexibility matters more than the provider’s legal structure. Its Eden AI catalog includes 24 models compared with Scaleway’s 17, offers context windows reaching 1 million tokens, starts at lower per-token pricing and supports deployment of customer-owned fine-tuned weights.

The practical decision is not simply which provider has European data centers. Choose Scaleway when jurisdiction, HDS certification, ANSSI alignment or French-hosted Mistral access determines procurement. Choose Nebius when you need a broader catalog, very long context windows, lower entry pricing or custom model weights.

Scaleway vs Mistral

Mistral is an AI lab that develops and maintains its own model family, while Scaleway is a cloud provider that hosts Mistral models alongside models from Qwen, Google, Meta, OpenAI’s open-weight range and H Company. A direct Mistral relationship provides first-party model support, greater consistency across the product roadmap and access to proprietary Mistral models that may not be available through third-party hosts.

Scaleway provides more vendor choice and reduces dependence on one lab’s release schedule. It also lets existing Scaleway customers run models such as Mistral Medium 3.5, Mistral Small 3.2, Devstral 2, Pixtral and Voxtral within the same French cloud account as their compute, storage, networking and Kubernetes workloads.

Choose Mistral when first-party support and access to its complete proprietary portfolio matter most. Choose Scaleway when you want Mistral coverage alongside alternative open-weight models and broader cloud infrastructure.

Scaleway vs OVHcloud

Scaleway and OVHcloud are both French cloud providers positioned for organizations evaluating sovereignty, jurisdiction and European infrastructure. OVHcloud has the larger infrastructure footprint and a longer-established position in European hosting, which can matter for organizations consolidating extensive compute, storage, networking and private-cloud estates under one supplier. It also brings its own sovereignty and compliance credentials to regulated procurement processes.

Scaleway’s advantage for AI inference is the breadth and recency of its serverless model catalog available through Eden AI. Its 17-model selection includes recent general-purpose, coding, multimodal and embedding models, plus less common specialized options such as scaleway/holo2-30b-a3b for computer use and scaleway/voxtral-small-24b-2507 for audio understanding.

Choose OVHcloud when infrastructure scale and consolidation are the primary requirements. Choose Scaleway when serverless model variety, recent releases and specialized AI capabilities are more important.

Similar providers available on Eden AI

• ‍Nebius: Best for broad open-weight model coverage, 1M-token context options, low entry pricing and deployment of customer-owned fine-tuned weights.‍
• Mistral: Best for first-party access to Mistral models, direct model support and consistency with the lab’s own release roadmap.‍
• OVHcloud: Best for organizations prioritizing a large French cloud footprint, infrastructure consolidation and established European sovereignty credentials.

Frequently asked questions about Scaleway on Eden AI