Mistral Saba
Mistral Saba, a 24B regional model for Arabic and South Asian languages, runs on a single GPU at over 150 tokens per second.
A language-and-culture specialist for the Middle East and South Asia, with notable strength in Tamil and other South Indian languages. Mistral says it beats models over 5x its size on those languages.
- Date
- Monday, 17 February 2025
- Lab
- Mistral AI
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | 24B single-GPU, >150 tokens/s | company |
Sources
This record was checked against its sources on 6 October 2026. How we check