Skip to content
llmwaves

Saba

Mistral AI · released Feb 17, 2025

ProprietaryTool useStructured output

Mistral Saba is a 24B-parameter language model specifically designed for the Middle East and South Asia, delivering accurate and contextually relevant responses while maintaining efficient performance. Trained on curated regional...

Specification

Context window
33K
Max output
Knowledge cutoff
Sep 30, 2024
Parameters
Undisclosed
Licence
Proprietary
Serving providers
1
Moderated
No

Intelligence

6.2

20th percentile

Coding

Coding Index

Agentic

Agentic Index

Output speed

62 t/s

Median across providers

Latency

445ms

Time to first token

Cost per task

$0.03

Estimated

Benchmarks

Where the score comes from

The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.

Evaluation scores

Percentage correct · higher is better

  • MMLU-Pro
    61.1%
  • GPQA Diamond
    42.4%
  • SciCode
    24.1%
  • AIME 2025
    13.0%
  • Humanity's Last Exam
    4.3%

An evaluation missing from this list was not run for this model — it is not a zero.

View as table
EvaluationScore
MMLU-Pro61.1%
GPQA Diamond42.4%
SciCode24.1%
AIME 202513.0%
Humanity's Last Exam4.3%

Against its peers

Intelligence Index · this model highlighted, nearest peers in grey

Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.

View as table
ModelIntelligence
GLM 4.5V6.8
GPT-46.8
Gemini 2.5 Flash Lite6.7
GPT-4o-mini6.7
GPT-4o-mini (2024-07-18)6.7
Granite 4.1 8B6.4
Saba6.2
Olmo 3 32B Think6.1

Percentile among all indexed models

Intelligence20th

Pricing

What it costs to run

List prices per million tokens, plus what one representative task works out to.

List price

Input / 1M tokens
$0.2
Output / 1M tokens
$0.6
Cached input / 1M
$0.02
Blended 3:1
$0.3

One task, estimated

$0.03

Input tokens
50,000
Output tokens
25,000
Profile
Standard

Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).

Serving providers

Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.