Mistral Large 4: Open Weights, Pricing, Licence and Mistral Studio

Last updated: 9 November 2026

Mistral Large 4 specifications

Mixture-of-Experts means only a fraction of the model is used for each token. That keeps running costs much lower than the total size suggests, which is why the API price is low. For background, see what is an LLM.

Pricing

Reasoning tokens count toward output, so High reasoning mode costs more per task than Low. Compared with leading closed models, this is among the cheapest frontier-class pricing available. See LLM API pricing comparison.

The licence: open, with conditions

Large 4’s weights are released under the Mistral Ecosystem Open-Weight License, described as a modified MIT licence. In plain terms:

• Internal use is allowed, including running it inside your company for your own products and staff.
• Using it through Mistral’s own products is outside the restriction.
• Model-as-a-service businesses, meaning companies that sell access to the model to others, need a separate commercial licence from Mistral before commercial use.

This is a common pattern for 2026 open-weight models: free for most users, with conditions for large hosting businesses. Kimi K3.1 Pro uses a similar approach. Read the licence yourself or ask a lawyer before building a product on it. See Kimi K3.1 Pro and open-source AI models.

Can you run Mistral Large 4 yourself?

Technically yes, practically only with data-centre hardware. Mistral’s model card deployment examples for vLLM, SGLang and Transformers all use eight GPUs with tensor parallelism. A separate NVFP4 checkpoint targets NVIDIA Blackwell GPUs such as B200 and GB200. Mistral recommends its API if local serving performs poorly.

For most businesses, the realistic options are:

• Mistral’s API or Le Chat. See how to use Mistral Le Chat.
• A cloud provider that hosts the model.
• Your own GPU cluster, if you already run one and have strict data control needs.

Smaller open models are a better fit for laptops and single GPUs. See how to run an LLM locally.

What Mistral reports about performance

These figures are Mistral’s own and have not all been independently confirmed.

Mistral attributes the gains to about three weeks of extra reinforcement learning and promises a full technical report. A specialised penetration-testing variant is available only in preview to vetted enterprise partners.

Mistral Studio

Alongside Large 4, Mistral launched Mistral Studio in early November 2026, a platform for building agents, workflows and internal apps on Large 4. Its main ideas:

• Workflows that combine fixed rules, agent reasoning and human approval steps.
• Traceable logs of what agents did and why.
• Flexible deployment: in the customer’s own cloud, a private network or on-premises.
• Named enterprise users include BNP Paribas, Capgemini and Accenture.

Mistral had not published Studio pricing at the time of writing. Studio’s pitch is aimed at banks, insurers, governments and other organisations that need control over where data goes and evidence of how decisions were made.

Who should consider Mistral Large 4

It may be less suitable if you need the very highest scores on the hardest reasoning tasks, or if you want to resell model access without a commercial agreement.

How it compares with other open models

See Kimi K3.1 Pro and DeepSeek API pricing for other low-cost options.

Getting started

1. Try it in Le Chat or the Mistral API with your real tasks.
2. Compare cost and quality against your current model on 20 to 50 representative prompts.
3. Test Low, Medium and High reasoning to find the cheapest setting that meets your quality bar.
4. Read the licence if you plan to self-host or build a product.
5. For agents in regulated settings, evaluate Mistral Studio’s approval and logging features.

Choosing a reasoning mode

Because High is the default, many teams overspend on simple tasks. Set the mode explicitly per task type. Low replaced the earlier “None” option from the preview.

Calling Mistral Large 4 from code

Mistral provides its own SDKs and a REST API. A minimal Python example looks like this; check Mistral’s documentation for the exact model identifier and parameters for Large 4:

from mistralai import Mistral

client = Mistral(api_key="YOUR_API_KEY")
resp = client.chat.complete(
    model="MISTRAL_LARGE_4_MODEL_ID",
    messages=[{"role": "user", "content": "Draft a GDPR-friendly privacy notice summary for a small online store."}],
)
print(resp.choices[0].message.content)

Store keys securely and set spending limits in your Mistral account.

Cost example

Suppose you summarise 1,000 documents a day, each with 8,000 input tokens and 800 output tokens, using Medium reasoning that adds roughly 400 reasoning tokens per request. At Large 4’s list prices:

• Input: 8 million tokens at $0.70 per million = $5.60 a day
• Output and reasoning: 1.2 million tokens at $2.10 per million = $2.52 a day
• Total: about $8.12 a day, or roughly $245 a month

If the same system prompt is reused, cached input at $0.07 per million can reduce the input cost further. Your real numbers will differ; measure on a sample first. See how to estimate AI API costs.

Data residency and compliance

For European organisations, Mistral’s appeal is partly about where data is processed and who controls the model. Options range from Mistral’s hosted API to self-hosting in your own data centre. Questions to settle with your legal and security teams:

• Where will prompts and outputs be processed and stored?
• Do you need on-premises deployment for regulatory reasons?
• What logging and audit evidence do you need for AI decisions?
• Does your use fall under the EU AI Act’s transparency rules? See EU AI Act Article 50.

Mistral’s model line-up in brief

Large 4 sits at the top of Mistral’s range. Mistral also released Voxtral 3 speech models in October 2026 and offers smaller models for lighter tasks. For consumer use, Le Chat gives access to Mistral’s models without code. See how to use Mistral Le Chat and ChatGPT alternatives.

Frequently asked questions