Mistral Small 3.1

// Overview

Mistral Small 3.1 is a 24-billion-parameter open-weight model released by Mistral AI on March 17, 2025, under the Apache 2.0 license. Carrying the API name mistral-small-2503, it upgrades the earlier Mistral Small 3 by adding state-of-the-art vision understanding and extending the context window to 128K tokens, all without sacrificing the text performance of the previous version.

Mistral positions Mistral Small 3.1 as the best model in its weight class, claiming it outperforms comparable small models such as Google's Gemma 3 and OpenAI's GPT-4o Mini while running at roughly 150 tokens per second. It ships with native function calling and JSON output for agentic use, strong multilingual coverage across dozens of languages, and a compact footprint that Mistral says can run on a single RTX 4090 or a 32GB-RAM Mac once quantized.

Both a base checkpoint and an instruction-tuned checkpoint are published on Hugging Face. Beyond Mistral's own La Plateforme API, Mistral Small 3.1 is available through Google Cloud Vertex AI and a range of third-party hosts including OpenRouter, NVIDIA NIM, and Ollama, making it a practical drop-in for self-hosted and cost-sensitive multimodal workloads.

// Benchmarks

1. MMLU80.62%
2. MMLU Pro (5-shot CoT)66.76%
3. GPQA Diamond (5-shot CoT)45.96%
4. MATH69.3%
5. HumanEval88.41%
6. MBPP74.71%
7. MMMU (vision)64%
8. MathVista68.91%
9. ChartQA86.24%
10. DocVQA94.08%
11. AI2D93.72%
12. RULER 32K (long context)93.96%
13. RULER 128K (long context)81.2%
14. Multilingual (average)71.18%

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

// Pricing

OpenRouter list price for mistral-small-3.1-24b-instruct. As an open-weight Apache-2.0 model it can also be self-hosted at no per-token cost.

Pricing source ↗

// Strengths

• Adds vision/image understanding to the 24B Small line while keeping text quality, with strong document and chart scores (DocVQA 94.1%, AI2D 93.7%, ChartQA 86.2%)
• Long 128K context with solid retrieval (RULER 32K 94.0%, RULER 128K 81.2%)
• Apache 2.0 license with both base and instruct weights on Hugging Face — fully self-hostable and fine-tunable
• Compact enough for single-GPU or 32GB-RAM local deployment, with ~150 tokens/sec inference
• Native function calling and JSON output for agentic and tool-using workflows
• Broad multilingual coverage (71.2% average across European, East Asian, and Middle Eastern language sets)

// Best for

• Self-hosted multimodal assistants that need to read images, documents, and charts
• On-device or single-GPU deployment where data privacy or latency rules out cloud APIs
• Document understanding and visual Q&A (forms, receipts, diagrams)
• Agentic pipelines using native function calling and structured JSON output
• Multilingual chat and reasoning across European, East Asian, and Middle Eastern languages
• Domain fine-tuning (legal, medical, technical) on top of the open base checkpoint

// How to access

// 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.