News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgMistral Medium 3

Mistral Medium 3

A mid-size language model aimed at enterprise coding and multimodal work

Mistral Medium 3 is a language model from Mistral AI, announced in 2025 as the third entry in the Mistral Medium line. The company positions it as a cost-efficient option for professional tasks such as coding and multimodal understanding, with deployment on cloud or in private infrastructure.

Share on WhatsApp

  • Announced by Mistral AI in 2025 as part of the Mistral Medium line
  • Company claims 90% or more of Claude Sonnet 3.7 benchmark performance
  • Priced at $0.4 per million input tokens and $2 per million output tokens
  • Can be self-hosted on four GPUs or more, or run in a private cloud
  • Fine-tuning and continued pretraining offered for enterprise domains

What it is

Mistral Medium 3 is a language model announced by Mistral AI in 2025. It sits in the company's Mistral Medium line, between its smaller Mistral Small and Ministral models and its larger Mistral Large offerings. Mistral AI describes it as a mid-size model built for professional work, especially coding and understanding of mixed text and image inputs. The company says it reaches close to the level of much larger and slower rivals on these tasks while costing far less to run. It is offered through an API and can also be deployed by customers on their own infrastructure.

How it came about

Mistral AI has released a series of models since Mistral 7B, including Mistral Small, Mistral Large, Pixtral, Mistral OCR, Mistral Saba and the Ministral 3B and 8B models. Mistral Medium 3 followed the Mistral Small launch in March of the same year. The company framed the release as a step towards a larger model it said was coming in the following weeks, and hinted that this future model would be open.

Inside the model

Mistral AI has not published architectural details such as layer counts, parameter totals or training data for Mistral Medium 3. The company says the model can be continuously pretrained and fully fine-tuned, and blended into enterprise knowledge bases, so that it can be adapted to specific domains and updated over time. It is designed to run in hybrid setups, on premises or inside a customer's virtual private cloud, and the company states it can be self-hosted on four GPUs or more.

What it can do

Mistral AI reports that Mistral Medium 3 performs at or above 90% of Claude Sonnet 3.7 across benchmarks, at $0.4 per million input tokens and $2 per million output tokens. The company says it also exceeds Llama 4 Maverick and Cohere Command A, and undercuts DeepSeek v3 on price in both API and self-deployed use. It is said to be strongest in coding and STEM tasks, and third-party human evaluations are cited as backing its coding performance against larger competitors.

How to use it

The model is served through Mistral La Plateforme and Amazon SageMaker, with IBM watsonx, NVIDIA NIM, Azure AI Foundry and Google Cloud Vertex listed as coming later. Customers wanting to run it in their own environment are asked to contact Mistral AI. The company offers continuous pretraining, full fine-tuning and integration with enterprise knowledge bases through its applied AI team.

Reception and debate

Beta customers in financial services, energy and healthcare are described by Mistral AI as using the model for customer service with deep context, personalised business processes and analysis of complex datasets. No independent reviews, criticisms or incidents are reported here.

Why it matters

Mistral AI presents Mistral Medium 3 as evidence that a medium-sized model can match flagship open models such as Llama 4 Maverick while being cheaper to serve. The release sits between the Mistral Small launch in March and a larger model the company signalled for the following weeks.

Questions people ask about Mistral Medium 3

What is Mistral Medium 3 for?

Mistral AI built it for professional use, with coding and multimodal understanding as the main focus. The company also targets enterprise work such as customer service, business process personalisation and analysis of large or complex datasets in sectors like finance, energy and healthcare.

How much does it cost to run?

The company lists API pricing at $0.4 per million input tokens and $2 per million output tokens. It says this is well below comparable models, and cheaper than DeepSeek v3 both through the API and when customers deploy the model themselves.

Can I run it myself?

Yes. Mistral AI says the model can be deployed on any cloud or in a customer's own environment, including self-hosted setups with four GPUs or more. Hybrid and in-VPC deployment are supported, and the company asks interested customers to get in touch for on-premises use.

How does it compare to other models?

Mistral AI reports performance at or above 90% of Claude Sonnet 3.7 across benchmarks, and says it beats Llama 4 Maverick and Cohere Command A. These are the company's own figures, measured through its internal evaluation pipeline, with third-party human evaluations cited for coding.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Mistral Medium 3 yet, so the facts on this page come from the sources listed above: mistral.ai/news/mistral-medium-3. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.