Mistral Large 2 (24.07)

// Overview

Mistral Large 2 (version 24.07, API name mistral-large-2407) is the second-generation flagship large language model from French AI lab Mistral AI, announced on July 24, 2024 under the banner "Large Enough." It is a dense transformer with 123 billion parameters and a 128,000-token context window, released with open weights on Hugging Face as Mistral-Large-Instruct-2407. Mistral positioned it as a single model that performs on par with leading frontier systems of its era such as GPT-4o, Claude 3 Opus, and Llama 3.1 405B, while being far smaller and cheaper to run.

The model was built for code, math, and reasoning. Mistral Large 2 supports a dozen major human languages — including English, French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean — plus more than 80 programming languages such as Python, Java, C, C++, JavaScript, and Bash. It also ships with native function calling, capable of both parallel and sequential tool execution, which made it a practical engine for agentic and retrieval-augmented applications.

Mistral Large 2 was distributed under the Mistral Research License, which permits research and non-commercial use of the open weights, with commercial deployment requiring a separate paid license. Through Mistral's hosted API (La Plateforme) and partner clouds it was a closed, paid endpoint. It has since been retired by Mistral — with a stated retirement date of March 30, 2025 — and superseded first by Mistral Large 24.11 (2.1) and then by Mistral Large 3, but the open weights remain available for self-hosting.

// Benchmarks

1. MMLU (pretrained, accuracy)84%

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

// Pricing

Historical hosted-API price for mistral-large-2407 (Mistral La Plateforme and partner platforms). The model is now retired and no longer served on Mistral's API; open weights remain free to self-host under the Mistral Research License for non-commercial use.

Pricing source ↗

// Strengths

• Strong code generation and reasoning for its size — Mistral pitched it as rivaling GPT-4o and Llama 3.1 405B at roughly a quarter of the latter's parameter count.
• Genuinely multilingual: 12 major languages plus 80+ programming languages, with notable gains over the original Mistral Large on non-English text.
• Open weights under the Mistral Research License, downloadable from Hugging Face for self-hosting and research.
• Native function calling with parallel and sequential tool execution, suited to agentic and RAG pipelines.
• Large 128k-token context window for long documents and multi-turn conversations.
• Mistral tuned it to reduce hallucination and to produce more concise, business-ready answers rather than padding.

// Best for

• Multilingual chat assistants and customer support across European and Asian languages.
• Code generation, completion, and review across many programming languages.
• Agentic workflows and tool use that rely on native function calling.
• Retrieval-augmented generation over long documents within the 128k context.
• Research and experimentation on a self-hosted open-weight frontier-class model (under the research license).
• Math and structured-reasoning tasks where the original Mistral Large fell short.

// How to access

// Mistral Large — every version

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