Command Palette

Search for a command to run...

Mistral Saba

Mistral

Proprietary

Context

Release Date
Feb 17, 2025
Window
32,000

PricingPer 1M tokens

Input
$0.2
Output
$0.6
Blended 3:1
$0.3

Capabilities

Speed
86 t/s
Input
Output
Reasoning tokens

Latency

TTFT
0.29 ms
500 token response
6.12 s

Benchmarks

Intelligence
●●○○○
Math
●●○○○
MMLU Pro
61.1%
GPQA
42.4%
HLE
4.1%
SciCode
24.1%
AIME
13.0%
MATH 500
67.7%
HumanEval
85.4%

Mistral Saba is a 24 billion-parameter language model tuned for Arabic and several Indian-origin languages, especially Tamil. It delivers higher accuracy than much larger general models, streams ~150 tokens/s, and runs on a single GPU—available both as a cloud API and for fully local, on-prem deployment.

Developers get a fast, low-cost way to build chatbots, content generators, or domain-specific tools that sound native to Middle-Eastern and South-Asian users. The model’s permissive deployment, fine-tuning support, and regional nuance mean you can ship compliant, high-quality features without juggling oversized models or external data-privacy risks.