Generative APIs API

Scaleway Generative APIs provides access to the latest AI models hosted on Scaleway infrastructure.

Generative APIs specification targets OpenAI API compatibility.

(switchcolumn) (switchcolumn)

Concepts

Refer to our dedicated concepts pageOpen in new context to find definitions of the different terms referring to Generative APIs.

(switchcolumn) (switchcolumn)

Quickstart

1. Configure your environment variables.

   Note

   This is an optional step that seeks to simplify your usage of the APIs.

   Code
   export SCW_ACCESS_KEY="<API access key>"
   export SCW_SECRET_KEY="<API secret key>"
   export SCW_REGION="<Scaleway region>"

2. Generate content from a model by running the following command.

   Code
   curl https://api.scaleway.ai/v1/chat/completions \
     -H "Content-Type: application/json" \
     -H "Authorization: Bearer $SCW_SECRET_KEY" \
     -d '{
       "model": "llama-3.3-70b-instruct",
       "messages": [
         {
           "role": "system",
           "content": "You are a helpful assistant."
         },
         {
           "role": "user",
           "content": "Hello!"
         }
       ]
     }'

See How to use Generative APIsOpen in new context for quickstart and snippets using REST request or libraries such as openai python client.

(switchcolumn)

• you have an account and are logged into the Scaleway consoleOpen in new context
• you have created an API keyOpen in new context and that the API key has sufficient IAM permissionsOpen in new context to perform the actions described on this page.
• you have installed curlOpen in new context

(switchcolumn)

Technical Information

Regions

Scaleway's infrastructure is spread across different regions and Availability ZonesOpen in new context.

Generative APIs is available in the Paris region, which is represented by the following path parameters (optional while there is only one region):

• fr-par

Supported endpoints and features

Supported endpoints are:

• /v1/responses
• /v1/chat/completions
• /v1/audio/transcriptions
• /v1/embeddings
• /v1/rerank
• /v1/batches
• /v1/models

The /v1/chat/completions endpoint:

• Supports many features such as:
  • Structured outputs (JSON response format)
  • Tool calling (ie. compatibility with workflows using MCP servers)
  • Sending and analyzing images
  • Sending and analyzing audio files
• Does not yet support the following parameters: audio, metadata, modalities, prediction, prompt_cache_key, user, safety_identifier, service_tier, stream_options.include_obfuscation, store, system_fingerprint, web_search_options.
• Does not support custom tools (only function tools are supported). These tools require you to provide a custom grammar to verify their input format.

The /v1/audio/transcriptions endpoint:

• Does not yet support the following parameters: chunking_strategy, include[] and timestamp_granularities[].

The /v1/responses endpoint:

• Does not yet support storing conversation state server-side.
• Does not support execution by Scaleway of "built-in" tools (such as web or file search) while the model generates a response. Currently, only function tools are supported. These tools must be fully defined in your query, and should be executed if the model requests them. The execution result is then sent to the model so that it can use this additional context to produce a suitable answer.
• Does not support sending response output object directly as an input for the next message. Standard messages using role:assistant and content fields should be used instead. Specifically, custom_tool_call and custom_tool_call_output types are not yet supported. As a workaround, tool call results can be send using role:user although we recommend using /chat/completions for better results in this case.
• Does not yet support the following parameters: background, conversation, include, instructions, max_tool_calls, metadata, previous_response_id, prompt_cache_key, safety_identifier, service_tier, stream_options, top_logprobs, user, prompt, verbosity.

The /v1/batches endpoint:

• Supports processing files stored in Object Storage (using Amazon S3 protocol).
• Does not yet support the following parameters: metadata.

The /v1/rerank endpoint:

• Aims for compatibility with JinAI API and Cohere API format (no OpenAI API exists for this endpoint)

Third party tool integration

For full details of direct integration into third party tooling, see Integrating Scaleway Generative APIs with popular AI toolsOpen in new context. If your tool is not listed, you can still specify the Scaleway URL and API key in most OpenAI-like plugins, as compatibility largely depends on the above APIs.

Technical Limitations

When choosing a model, select the ones compatible with the APIs endpoints you want to use in our model catalogOpen in new context. For example, /v1/embeddings is only available for embeddings models.

Going Further

For more information about Generative APIs, you can check out the following pages:

• Generative APIs DocumentationOpen in new context
• Generative APIs FAQOpen in new context
• Scaleway Slack Community - AI ChannelOpen in new context join the #ai channel
• Contact our support teamOpen in new context

Troubleshoooting

See Troubleshooting Generative APIsOpen in new context for advanced APIs behaviour descriptions and common issues solutions.