ai/magistral-small-3.2

Verified Publisher

By Docker

Updated 10 months ago

24B multimodal instruction model by Mistral AI, tuned for accuracy, tool use & fewer repeats

Model
1

10K+

ai/magistral-small-3.2 repository overview

Mistral-Small-3.2-24B-Instruct-2506

GGUF version by Unsloth

logo

Description

A 24B-parameter instruction-tuned multimodal model by Mistral AI, optimized for following instructions, reducing repetition errors, and supporting function/tool calling.

Characteristics

AttributeDetails
ProviderMistral AI
ArchitectureLlama
Cutoff date2023-10-01
Languages24 languages
Tool calling
Input modalitiesText, Images
Output modalitiesText
LicenseApache 2.0

Available model variants

Model variantParametersQuantizationContext windowVRAM¹Size
ai/magistral-small-3.2:24B

ai/magistral-small-3.2:24B-UD-Q4_K_XL

ai/magistral-small-3.2:latest
23.57 BMOSTLY_Q4_K_M41K tokens14.59 GiB13.50 GB
ai/magistral-small-3.2:24B-UD-IQ2_XXS23.57 BMOSTLY_IQ2_XXS41K tokens7.45 GiB6.28 GB

¹: VRAM estimated based on model characteristics.

latest24B

Use this AI model with Docker Model Runner

docker model run ai/magistral-small-2506

Considerations

  • We recommend using a relatively low temperature, such as temperature=0.15.
  • Make sure to add a system prompt to the model to best tailor it for your needs. If you want to use the model as a general assistant, we recommend to use the one provided in the SYSTEM_PROMPT.txt file.

Benchmark performance

CategoryMetricMistral-Small-3.2-24B-Instruct-2506
Instruction FollowingWildBench v265.33%
Arena Hard v243.1%
IF Accuracy (Internal)84.78%
STEMMMLU80.50%
MMLU Pro (5-shot CoT)69.06%
MATH69.42%
GPQA Diamond (5-shot CoT)46.13%
MBPP Plus Pass@578.33%
HumanEval Plus Pass@592.90%
SimpleQA (TotalAcc)12.10%
VisionMMMU62.5%
ChartQA87.4%
DocVQA94.86%
AI2D92.91%

Tag summary

Content type

Model

Digest

sha256:9c3bc0d23

Size

14.3 GB

Last updated

10 months ago

docker model pull ai/magistral-small-3.2:24B-UD-Q4_K_XL

This week's pulls

Pulls:

108

Last week