Mistral 3 Models: Sizes, Apache 2.0 License, and Pricing

Four models, and a reversed license choice

Mistral AI released four new model weights on December 2, 2025: Mistral Large 3, a mixture of experts model with 675 billion total parameters and 41 billion active per token, and three Ministral models at 14 billion, 8 billion and 3 billion parameters. Mistral published base and instruction fine-tuned versions of all four under the Apache 2.0 license, plus a reasoning variant for each of the three Ministral sizes; a reasoning version of Large 3 was described as coming later. Mistral said Large 3 was trained on 3,000 Nvidia H200 GPUs. [2]

Atlas interpretation: The blanket Apache 2.0 license reverses a path Mistral had already walked in the other direction. Its first open family, Mistral 7B, shipped fully open in 2023. Its code model, Codestral, shipped in 2024 under a license that barred commercial use without a separate deal, the moment the company that built its name on open weights drew a line. Mistral 3 puts the entire flagship line, including the largest model, back on the open side of that line. [2]

A leaderboard position Mistral reported itself

Mistral said Large 3 debuted at number two among open, non-reasoning models on the LMArena leaderboard, and sixth among all open models on the same leaderboard. Mistral also reported that Ministral 14B's reasoning variant scored 85 percent on the AIME 2025 math competition benchmark, and described the Ministral series as offering the best cost-to-performance ratio of any open-weight model, without naming the comparison set behind that claim. [2]

Atlas interpretation: Every figure in that paragraph comes from Mistral's own announcement. This page does not independently confirm the LMArena placement against the public leaderboard or rerun the AIME benchmark, so the ranking and the cost-to-performance claim should be read as self-reported until an independent source checks them. [2]

A vision model that runs in a browser tab

Simon Willison, testing the release the same day, called Ministral 3B a competent vision-capable model in a roughly 3 gigabyte file and ran it entirely on-device in a browser tab over WebGPU, with a live demo that streamed webcam input into the model and produced no server request after the page loaded. [3]

Willison also ran his standard pelican-riding-a-bicycle drawing test across all four models. Large 3 produced the most recognizable result of the family; Ministral 3B produced abstract shapes rather than a distinguishable bicycle or pelican. He published the images but did not attach a numeric score to either result, and did not run the same test against a competing vendor's models for comparison. [3]

Multi-cloud on day one, priced like a frontier model

Mistral made the family available immediately through its own Mistral AI Studio and through Amazon Bedrock, Microsoft Foundry, Hugging Face, Modal, IBM watsonx, OpenRouter, Fireworks, Unsloth AI and Together AI, with Nvidia NIM and AWS SageMaker listed as coming later. [2]

Microsoft's Foundry listing put Large 3 in public preview in the West US 3 region on the day of release, priced at 0.50 dollars per million input tokens and 1.50 dollars per million output tokens, rates typical of a hosted frontier model rather than a discount for openly licensed weights. [4]

Atlas interpretation: An Apache 2.0 license does not obligate Mistral to price hosted access cheaply, and charging frontier-model rates on a partner cloud for an openly licensed model is a choice about where the revenue comes from: not the weights, which anyone can download and run for the cost of the hardware, but managed inference for customers who would rather not operate 675 billion parameters themselves. [4]