Mistral Medium 3

// Overview

Mistral Medium 3 (model ID mistral-medium-2505) is a mid-tier large language model released by Mistral AI on May 7, 2025 under the tagline "Medium is the new large." Unlike many of Mistral's earlier models, Mistral Medium 3 is proprietary rather than open-weight — it is served through Mistral's API and cloud partners, and was positioned as a cost-efficient enterprise model that delivers frontier-level quality at roughly an order of magnitude lower cost than traditional large models.

Mistral AI claimed Mistral Medium 3 performs at or above 90% of Anthropic's Claude 3.7 Sonnet across benchmarks, while standing out in coding and STEM tasks and surpassing leading open models such as Llama 4 Maverick and enterprise models such as Cohere Command A. The model is multimodal: it accepts text and image input, handles document understanding (with OCR and strong DocVQA results), and supports a 128K-token (131K) context window.

Mistral Medium 3 was designed for flexible deployment — it can run on any cloud or in self-hosted environments of four GPUs and above. It launched on Mistral's La Plateforme and Amazon SageMaker, with availability on IBM watsonx, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex following. It was later succeeded by Mistral Medium 3.1 (August 2025) and Mistral Medium 3.5 (April 2026), and is being deprecated in 2026.

// Benchmarks

1. HumanEval (0-shot)92.1%
2. Math500 Instruct (0-shot)91%
3. MMLU Pro (5-shot CoT)77.2%
4. GPQA Diamond (5-shot CoT)57.1%
5. ArenaHard (0-shot)97.1%
6. IFEval (0-shot)89.4%
7. MMMU (0-shot)66.1%
8. DocVQA (0-shot)95.3%
9. RULER 128K90.2%
10. LiveCodeBench v6 (0-shot)30.3%

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

// Pricing

Pricing source ↗

// Strengths

• Strong coding performance — HumanEval 92.1%, close to much larger and slower competitors
• Competitive STEM and reasoning — Math500 91.0% and MMLU Pro 77.2%
• Multimodal: handles images and documents (DocVQA 95.3%, MMMU 66.1%)
• Very low cost for its tier at $0.40 input / $2.00 output per 1M tokens
• Flexible deployment — any cloud, self-hosted, or on-prem from four GPUs up
• Reliable instruction following (IFEval 89.4%) and long-context retrieval (RULER 128K 90.2%)

// Best for

• Cost-sensitive enterprise coding assistants and code generation
• STEM, math, and technical reasoning workloads
• Document and image understanding (OCR, DocVQA-style extraction)
• Customer-facing assistants that need quality at controlled cost
• On-premise or self-hosted deployments where data must stay in-house
• Long-context tasks up to 128K tokens

// How to access

// Mistral Medium — every version

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