Mistral Large

Mistral Large is the flagship general-purpose model family from French AI lab Mistral AI. The current generation, Mistral Large 3, is an Apache 2.0 mixture-of-experts model with a 256k context window and multimodal input, priced at $0.50 per million input tokens - roughly an order of magnitude below comparable US frontier APIs - and available as downloadable weights for fully on-premise deployment.

Mistral Large is the flagship general-purpose model line from French AI lab Mistral AI, and the name now refers to Mistral Large 3 (API name mistral-large-3-25-12), released on 2 December 2025. Large 3 is a granular mixture-of-experts model with 675 billion total parameters and 41 billion active per token, a 256,000-token context window, and native multimodal input: it accepts text and images and returns text. Critically for enterprise buyers, Mistral shipped it under Apache 2.0 - a genuinely permissive open-source licence with no monthly-active-user ceiling and no separate commercial agreement. That is a sharp break from its predecessor. Mistral Large 2 (123 billion dense parameters, 128k context, released 24 July 2024, 84.0% MMLU, over 80 programming languages, parallel and sequential function calling) was released under the Mistral Research License, which required a paid commercial agreement for any self-deployed business use; Artificial Analysis now marks Large 2 deprecated in favour of Large 3. On Mistral's own API, Large 3 lists at $0.50 per million input tokens and $1.50 per million output tokens - roughly 80% and 90% below GPT-5.4 on input and output respectively by CloudZero's comparison, and 83%/90% below Claude Sonnet 4.6. It supports function calling, structured outputs, document Q&A, batching, conversations and built-in tools. Distribution runs through la Plateforme, Le Chat, and the major cloud marketplaces including Azure AI, Amazon Bedrock, Google Cloud and IBM watsonx.ai; because the weights are open it can also run entirely on-premise, which is the reason European banks and regulated buyers shortlist it over US-hosted-only alternatives.

Category

AI Models & APIs

Pricing

Usage-based, Subscription, Freemium, Contact for pricing

Target Market

CIOs, CTOs, Enterprise Developers, Data Protection Officers, Heads of AI

Deployment

API-based, Open-source, Self-hosted, Multi-cloud, Hybrid

Founded

2023

Headquarters

Paris, France

Team Size

201-500

• ✓
  Granular mixture-of-experts architecture

  675B total parameters with only 41B active per token, so serving cost tracks the small number while capability tracks the large one.

• ✓
  256,000-token context window

  Handles book-length contracts, full repositories or large retrieval sets in one call without an external chunking pipeline.

• ✓
  Apache 2.0 open weights

  No MAU ceiling and no commercial agreement required, which is materially more permissive than the community licences on competing open-weight models.

• ✓
  Native multimodal input

  Accepts images alongside text, so document, diagram and screenshot workflows do not need a separate vision model.

• ✓
  Function calling and structured outputs

  Supports parallel and sequential tool calls with schema-constrained JSON, which is the prerequisite for reliable agent orchestration.

• ✓
  Multi-cloud and on-premise distribution

  Available on la Plateforme, Azure AI, Amazon Bedrock, Google Cloud and IBM watsonx.ai, or self-deployed from the open weights.

• •
  Sovereign enterprise assistant

  European banks and insurers deploy the open weights in-country so no customer data crosses a jurisdictional boundary during inference.

• •
  High-volume document extraction

  Process contracts, claims and invoices at $0.50 per million input tokens where a frontier US API would cost roughly ten times more.

• •
  Multilingual customer operations

  Handle French, German, Spanish, Italian and Portuguese support queues on one model rather than one vendor per language.

• •
  Tool-calling agent backends

  Drive internal automations using parallel function calls and structured outputs, with the 256k window holding full tool schemas and history.

• •
  Code assistance across large repositories

  The Large 2 lineage trained on 80+ programming languages, and the long context lets a whole service be reasoned about in one prompt.

Mistral publishes list API pricing, which is unusual at this tier: Large 3 is metered at $0.50 per million input tokens and $1.50 per million output, which CloudZero measures at roughly 80% cheaper on input and 90% cheaper on output than GPT-5.4, and 83%/90% cheaper than Claude Sonnet 4.6. The Apache 2.0 weights make the per-token cost avoidable entirely if you have the GPUs, but 675B total parameters means that is a cluster decision rather than a workstation one. Two billing traps matter: Le Chat subscriptions (Free, Pro $14.99/mo, Team $24.99/user/mo) are a completely separate stream from API credits and buy no API access, and Le Chat Enterprise with SAML SSO and white-labelling is quoted rather than listed.

This page was written from 5 sources, 4 on domains other than mistral.ai.

1. 1.mistral.ai — mistral large 2407vendor
2. 2.docs.mistral.ai — mistral large 3 25 12
3. 3.artificialanalysis.ai — mistral large 2
4. 4.cloudzero.com — mistral api pricing
5. 5.openrouter.ai — mistral large