Skip to content
What is Mistral Large 4? Mistral's open-weight 1T model
Back to Blog
#AI models#Mistral Large 4#Mistral Large 4 pricing#Mistral Large 4 API#open-weight model#Mistral AI

What is Mistral Large 4? Mistral's open-weight 1T model

Sadaf Masood

Mistral Large 4 is Mistral's largest model so far: a multimodal mixture-of-experts model with about 1 trillion parameters, 52 billion of them active per token. It launched on 6 October 2026 as a public preview you can call through Mistral's API today, at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the open weights follow by the end of October, so you will also be able to run it on your own servers.

If you are deciding whether to put it in a product, the short version: it is worth testing now for security work, document and image reading, and jobs that must stay in the EU. Treat it as a preview, not a finished model.

"You can try the preview API today on Mistral Studio. Weights drop end of this month."

— Mistral, Introducing Mistral Large 4

What can it do?

It reads text and images and handles long inputs. Mistral's docs list a 1M-token context window, a 1.6B-parameter vision encoder and support for function calling, structured outputs, document Q&A, batch jobs and Mistral's agents API. The API model id is mistral-large-4.

Mistral pitches it at four kinds of work. All the scores below are Mistral's own published results, not independent tests:

  • Security work. 93% of the 40 Cybench challenges, and 82% on a test that asks a model to reproduce a real vulnerability and then patch it. Mistral says several closed models score near zero on that test because they refuse it.
  • Coding agents. 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. In a blind human rating of code quality, it came second of five, behind Claude Opus 5.
  • Business workflows. 59.9% on AutomationBench, which covers 657 workflows across tools like Gmail, Sheets, Slack and Salesforce.
  • Reading documents and images. Charts, PDFs, engineering drawings and satellite images. Mistral reports it slightly ahead of GPT-6 Astra on one visual grounding test (42% vs 41%).

It was trained on more than 160 languages, including every official EU language, and resisted 93.3% of attacks on Lakera's prompt-injection benchmark.

What does it cost?

List price is $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens (Mistral's model page, read 11 October 2026). The same page currently shows a sale price of half that: $0.68 input, $0.07 cached and $2.09 output. It does not say when the sale ends, so budget on the list price.

A worked example. You extract fields from 10,000 contracts or invoices, each about 3,000 tokens in and 500 tokens out:

  • Input: 10,000 × 3,000 = 30M tokens × $1.36 = $40.80
  • Output: 10,000 × 500 = 5M tokens × $4.18 = $20.90
  • Total at list price: $61.70. At today's sale price: $20.40 + $10.45 = $30.85.

For context, Mistral's pricing page gives its previous Mistral Large at $0.50 in and $1.50 out. Large 4 costs about 2.7 times as much per token at list price. Test it on your own task before you move high-volume work across.

Should you build on it now?

Test it now. Wait before you lock it into a production product. Three facts decide this:

  • It is still being trained. Mistral says the reinforcement-learning run behind the preview is "still in flight" and expects big improvements in the coming weeks. That is good news, but outputs may change under you. Keep a fixed set of test cases and run them again when Mistral updates the model.
  • Self-hosting is not available yet. If you need the model on your own servers or private cloud, for regulated data or offline use, that depends on the weights shipping as promised. Check the licence terms when they do.
  • Where the data goes. The preview runs in Mistral's own European datacentres. Mistral says it will also be available in other regions, including a European deployment it runs under EU law. For buyers who need EU data residency, this is the main reason to choose it over a US model.

A view, not a fact: it is a strong first choice for security tooling, multilingual document work and EU-hosted products. If your job is cheap, high-volume classification, a small model will cost far less.

Want it in your product?

If you want Mistral Large 4 built into your product or workflow, with a test set and cost limits from day one, see what we build or tell us about the job.

Sources