Mistral Saba
2025-02-17 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI introduces Mistral Saba, its first regional model: 24 billion parameters trained on data from the Middle East and South Asia, strong in Arabic and Indian languages such as Tamil.
What changes
- Mistral's first regional language model, trained on carefully curated datasets from these regions.
- Support for Arabic and many Indian languages, with a particular strength in South Indian languages (Tamil).
- According to the vendor, more accurate answers than models more than 5 times larger, at lower cost and latency.
- Deployable locally on the customer's premises, on a single GPU, at more than 150 tokens/s.
### Specifications - 24 billion parameters; 32k token context (documentation). - Text only. - License: proprietary (Premier model), weights not published; local deployment for enterprise customers. - Availability: API on La Plateforme, on-premises deployment on commercial request. - API identifier: `mistral-saba-2502` (alias `mistral-saba-latest`, retired on 09/30/2025 according to the documentation). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Mistral Saba | 0.2 | 0.6 | Price recorded from the pricing page archived on June 20, 2025; not published in the announcement post. ### Benchmarks No figures in text in the announcement (the comparison charts are images). The vendor claims more accurate answers than models more than 5 times larger. ### Key takeaways An option for use cases in Arabic and Indian languages, with the possibility of on-premises deployment for organizations subject to sovereignty constraints.
Links
Tags : LLM, Mistral, Saba, Multilingue, Arabe