Curiosity · AI Model

Mistral Small 24B

Mistral Small 24B, released in January 2025, was Mistral's answer to the wave of 20-30B open-weights models (Gemma 2 27B, Qwen2.5 32B). At 24B parameters and 32K context, it fits on a single H100 and delivers competitive reasoning and coding scores, making it the default 'good-enough, on-prem' open model for many teams.

Model specs

Vendor
Mistral AI
Family
Mistral Small
Released
2025-01
Context window
32,768 tokens
Modalities
text
Input price
$0.2/M tok
Output price
$0.6/M tok
Pricing as of
2026-04-20

Strengths

  • Open weights under the Apache 2.0 license
  • Fits on a single H100 with high throughput
  • Strong reasoning and coding per parameter

Limitations

  • Below frontier closed models (Opus, GPT-5) on hard reasoning
  • Text-only — no vision or audio modalities in this variant
  • English-biased despite multilingual pretraining

Use cases

  • On-prem mid-size LLM deployments on single H100/A100
  • Fine-tuning targets for domain chatbots
  • Low-cost reasoning tier in tiered LLM pipelines
  • Retrieval-augmented Q&A over internal corpora

Benchmarks

BenchmarkScoreAs of
MMLU≈81%2025-01
HumanEval≈86%2025-01
MATH≈70%2025-01

Frequently asked questions

What is Mistral Small 24B?

Mistral Small 24B is Mistral AI's 24-billion-parameter open-weights dense decoder, released in January 2025 under Apache 2.0 and aimed at strong reasoning per dollar on single-GPU servers.

Where can I download Mistral Small 24B?

Weights are published on Hugging Face under 'mistralai/Mistral-Small-24B' and related instruct checkpoints, and Mistral hosts it on la Plateforme as a paid API model.

How does Mistral Small 24B compare to Qwen2.5 32B?

Both are strong open-weights models in the 20-30B range. Mistral Small 24B is lighter and Apache-licensed; Qwen2.5 32B is slightly stronger on math and coding but runs on a larger footprint.

Sources

  1. Mistral — Mistral Small 3 launch — accessed 2026-04-20
  2. Hugging Face — mistralai/Mistral-Small-24B — accessed 2026-04-20