Mistral Small 3.2

Mistral Small 3.2 is an open-weight 24-billion parameter instruction-following model that expands on its predecessor's capabilities while introducing stronger vision understanding and a longer context window. The model supports image understanding alongside text reasoning, making it suitable for document analysis and multimodal workflows. Its architecture balances capability with local-run practicality—Simon Willison noted these 24B parameter models hit a sweet spot between performance and laptop RAM consumption. The open weights status means developers can run it locally, deploy via Ollama, or serve through NVIDIA NIM containers depending on infrastructure needs.

The 3.2 release refines the Small 3.1 recipe with targeted post-training improvements: notably better instruction adherence, a two-fold reduction in infinite generation loops on repetitive prompts, and a more robust function calling template. Benchmark improvements are significant—Wildbench scores rose from 55.6% to 65.33% and Arena Hard v2 jumped from 19.56% to 43.1%. The model is recommended to run with a temperature of 0.15 rather than the more common 0.7 default, reflecting careful tuning of the inference guidance. Available on Hugging Face and NVIDIA NGC, it serves both local development and production deployment scenarios.

DevPass (LLM Gateway)mistral-small-2506mistral-small

Open provider page Provider docs

Quick Info

Provider: DevPass (LLM Gateway)
Model key: mistral-small-2506
Release date: Jun 20, 2025
Last updated: Jun 20, 2025
Knowledge cutoff: 2025-03
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $0.30

Limits

Output tokens: 16,384 tokens
Context window: 128,000 tokens

Transparent token rates

Compare Mistral Small 3.2 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Mistral Small 3.2

LLM Gateway

CoverageRelease Notes

Mistral AI launches Small 3.2 with improved instruction-following, reduced repetition, and enhanced function calling for seamless AI integration

DevPass (LLM Gateway)

CoverageBenchmark

Artificial Analysis's Mistral provider dashboard covers 19 models and explicitly names Mistral Small 3.2 with a blended price of $0.12 per 1M tokens, ranking it third lowest in pricing behind Ministral 3 3B ($0.10) and Mistral Small 3.1 ($0.12). In the output-speed ranking, Mistral Small 3.2 does not appear in the top The intelligence ranking is led by GLM-5.2 (max) at 34, followed by Mistral Medium 3.5 at 15 and Mistral Small 4 at 11, with Mistral Small 3.2 not featured in the top intelligence slots. Context-window leaders are GLM-5.2 (max) at 1M, Mistral Medium 3.5 at 262k, and Mistral Large 3 at 262k. The page is useful for posit

DevPass (LLM Gateway)

CoverageBenchmark

BenchmarkList hosts a dedicated evaluation page for Mistral Small 3.2 24B (mistral-small-3.2-24b-instruct) that positions the model against peers including Claude Opus 5, Kimi K3, Qwen3.8 variants, and GLM-5.3 entries. The page lists pricing at $0.075 per 1M input tokens and $0.20 per 1M output tokens, and reports a 44 Detail shown includes a 2026-06-28 self-reported Terminal-Bench Hard score and per-category breakdowns from the PersonaArena persona-level role-playing benchmark, with an average score of 3.75 and sub-scores across knowledge, behavior, engagement, persona, image, backstory, and adversarial dimensions. The page confirms

DevPass (LLM Gateway)

CoverageRelease Notes

Opper AI's dated Mistral release tracker lists Mistral Small 3.2 as a distinct entry under June 2025, alongside Magistral Medium 1 and Magistral Small 1. The catalog confirms the June 20, 2025 release date and treats Small 3.2 as a separate model from later Small 4 and Small 3.1 entries, with no separately reported con The same tracker shows the broader Mistral catalog reaching 31 total releases, with Mistral Medium 3.5 (April 29, 2026) as the most recent entry and Mistral Small 4 (March 16, 2026) as a successor in the same family. The page provides a useful chronological reference for placing Mistral Small 3.2 within Mistral's relea