AI/TLDR

Mistral Small 3.2

Mistral's 24B open-weight workhorse, tuned for instruction following and tool calls

Overview

Mistral Small 3.2 (full name Mistral-Small-3.2-24B-Instruct-2506, API id mistral-small-2506) is a 24-billion-parameter open-weight vision-language model released by Mistral AI on June 20, 2025 under the Apache 2.0 license. It is a maintenance-style refresh of Mistral Small 3.1: the underlying base model is unchanged, but the instruction tuning was reworked to follow precise prompts more reliably, repeat itself far less, and call tools more dependably.

The headline gains over 3.1 are practical rather than flashy. Mistral reports internal instruction-following accuracy rising to 84.78% from 82.75%, the rate of infinite or repetitive generations dropping from 2.11% to 1.29%, and large jumps on conversational benchmarks like Wildbench v2 (55.6% to 65.33%) and Arena Hard v2 (19.56% to 43.1%). Coding (HumanEval Plus, MBPP Plus) and most STEM and vision scores hold steady or improve slightly.

With a 128K-token context window, native image understanding, and an Apache 2.0 license, Mistral Small 3.2 is positioned as a self-hostable workhorse for teams that want a capable model they can run on their own GPU. Mistral lists the model as deprecated as of April 30, 2026, with Mistral Small 4 as the recommended successor, but the open weights remain freely downloadable on Hugging Face.

Released2025-06-20
LicenseApache 2.0
WeightsOpen weights
Parameters24B
Context128K
ArchitectureA dense 24-billion-parameter transformer built on the Mistral Small 3.1 24B base. Version 3.2 (2506) is an incremental instruction-tuning update rather than a new pretrain: it sharpens instruction following, roughly halves runaway repetition, and ships a more robust function-calling template. It is a vision-language model that accepts interleaved text and image inputs, and runs in bf16/fp16 on a single GPU with about 55 GB of memory (smaller quantized GGUF builds run on consumer hardware).
Knowledge cutoffOctober 2023
ModalitiesText, Vision
StatusDeprecated

Benchmarks

  1. HumanEval Plus (Pass@5, coding)92.9%
  2. MBPP Plus (Pass@5, coding)78.33%
  3. MMLU (5-shot)80.5%
  4. MMLU Pro (5-shot CoT)69.06%
  5. MATH69.42%
  6. GPQA Diamond (5-shot)46.13%
  7. Internal instruction-following accuracy84.78%
  8. Wildbench v265.33%
  9. Arena Hard v243.1%
  10. ChartQA (vision)87.4%
  11. DocVQA (vision)94.86%
  12. MMMU (vision)62.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Pricing

Input$0.10 per 1M tokens
Output$0.30 per 1M tokens

Pricing for the Mistral Small line on Mistral's La Plateforme API. Third-party hosts (e.g. OpenRouter, DeepInfra) may price the open weights differently; self-hosting the Apache 2.0 weights is free of license fees.

Pricing source ↗

Strengths

  • Apache 2.0 open weights you can download, fine-tune and self-host without usage restrictions
  • Stronger instruction following and roughly 2x fewer infinite or repetitive generations versus Mistral Small 3.1
  • More robust function-calling template, well suited to agent and tool-use pipelines (recommended with vLLM)
  • Native vision-language input: reasons over images, charts and documents alongside text
  • 128K-token context window for long documents and multi-turn sessions
  • Fits on a single ~55 GB GPU in bf16/fp16, with quantized GGUF builds for consumer hardware

Best for

  • Self-hosted chat assistants and internal copilots where data must stay on-premise
  • Agentic and tool-calling workflows that need reliable, structured function calls
  • Document and chart understanding (DocVQA, ChartQA-style tasks) from images plus text
  • Coding assistance and code generation at a low cost-per-token
  • Fine-tuning a permissively licensed base for domain-specific applications
  • Cost-sensitive, high-volume API workloads that don't need a frontier-scale model

How to access

ProviderModel ID
Mistral La Plateforme ↗mistral-small-2506
Hugging Face ↗mistralai/Mistral-Small-3.2-24B-Instruct-2506
Amazon Bedrock Marketplace / SageMaker JumpStart ↗Mistral-Small-3.2-24B-Instruct-2506
OpenRouter ↗mistralai/mistral-small-3.2-24b-instruct

Mistral Small — every version

The full lineage of the Mistral Small line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.

VersionReleasedContextLicense
Mistral Small 4current2026-03-16Apache-2.0
Mistral Small 3.22025-06-20Apache-2.0
Mistral Small 3.12025-03-17Open weights
Mistral Small 32025-01-30Apache-2.0
Mistral Small (24.09)2024-09-17Open weights

FAQ

Is Mistral Small 3.2 open source?

Yes. The weights are released on Hugging Face under the Apache 2.0 license, which permits commercial use, fine-tuning and self-hosting without usage restrictions or license fees.

What changed from Mistral Small 3.1 to 3.2?

Version 3.2 (2506) keeps the same 24B base but reworks the instruction tuning: better instruction following (internal accuracy 82.75% to 84.78%), roughly half as many infinite or repetitive generations (2.11% to 1.29%), and a more robust function-calling template. Most coding, STEM and vision scores hold or improve slightly.

Does Mistral Small 3.2 support images?

Yes. It is a vision-language model that accepts interleaved text and image inputs, with strong document and chart understanding (for example 94.86% on DocVQA and 87.4% on ChartQA). It does not handle audio, video or native PDF input.

Is Mistral Small 3.2 still recommended?

Mistral lists it as deprecated as of April 30, 2026, with Mistral Small 4 as the recommended successor on the API. The Apache 2.0 open weights, however, remain freely downloadable for self-hosting.