Mistral: Saba

Mistral: Saba is a text model from Mistral, released by the provider on 17 February 2025. It is a 24B-parameter language model specifically designed for the Middle East and South Asia, trained on curated regional data to deliver accurate and contextually relevant responses while maintaining efficient performance. Its knowledge cutoff is 30 September 2024.

The model accepts text and file inputs and returns text. It has a context window of 32,768 tokens and a maximum output of 26,214 tokens, so it can handle fairly long inputs and produce lengthy completions within a single request. Supported request parameters include frequency_penalty, max_tokens, presence_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools and top_p. The presence of tools and tool_choice means it can be used in function-calling or agent-style workflows, while structured_outputs and response_format let you constrain the shape of the returned text.

Because of its regional focus, Saba suits applications where Middle Eastern and South Asian language, terminology and context matter: customer support assistants, content generation and summarisation for audiences in those regions, document question answering over uploaded files, and extraction or classification tasks where locally relevant phrasing affects quality. The file input modality is useful for workflows that pass documents directly to the model rather than pasting text, and the tool parameters support pipelines that need to call external functions or APIs. The 32,768-token context is workable for long reports, multi-turn conversations and moderate document sets, though very large corpora will need chunking or retrieval.

You call the model through api14's OpenAI-compatible API using the model id "mistralai/mistral-saba". Point your existing OpenAI client at the api14 base URL, send a chat completions request with that id, and pass any of the supported parameters listed above in the usual way. Because the interface is OpenAI-compatible, code written for OpenAI models generally works with minimal changes beyond the model id and endpoint.

Keep the context and output limits in mind when designing requests: inputs beyond 32,768 tokens must be trimmed or split, and responses are capped at 26,214 tokens. If your task depends on knowledge after 30 September 2024, or falls outside the Middle East and South Asia focus where a general-purpose model may be a better fit, consider another model on api14 instead.