Mistral Large 4 Open Weights: What Changes for Businesses

Mistral Large 4 is the roughly 1-trillion-parameter AI model with a 1M-token context whose weights Mistral AI opened on October 29, 2026. Internal use, fine-tuning and commercial use are allowed; businesses that offer the model to others as a service need a separate license. What really changes for businesses is the option of data control. What does not change is the hardware: Mistral's reference setups use eight GPUs, so for most companies the API is still the right place to start.

What changed with Mistral Large 4 on October 29?

Mistral AI released the weights of Large 4 on October 29, after an API preview that began on October 6, 2026. By Mistral's own measurement, the Terminal-Bench 4.0 score rose from 28% in preview to 41%. API pricing per million tokens is $0.70 for input, $0.07 for cached input and $2.10 for output. Beri reports that preview list prices were $1.36 and $4.18, so the price roughly halved. According to the model card, Large 4 takes text and images as input, returns text, and supports a 1M-token context. It is also available through OpenRouter, NVIDIA NIM and Microsoft Foundry.

What does the Mistral Large 4 license allow?

The Mistral Ecosystem Open-Weight License is a modified MIT license that permits use, modification, fine-tuning and commercial use. The main restriction sits in its "Model as a Service" definition. Under the license text, if you give a third party access through an API or hosted endpoint in a way that lets them exercise meaningful control over inputs, parameters or training data, you need a separate commercial license from Mistral. Two cases are excluded: end-user products with model capabilities solely embedded within specific features, and merely relaying requests to models hosted by others.

Our reading: an assistant answering customer questions on your website most likely falls under the first exclusion, while a platform selling model access where customers write their own instructions needs the separate license. This is not legal advice. If your product sits between those two, review the license with your counsel.

What did not change with Mistral Large 4?

Running Mistral Large 4 on your own servers is still a data center job. The vLLM and SGLang examples in the model card target an eight-GPU node. Beri puts the FP8 checkpoint at about 1.05 TB, running on an eight-GPU NVIDIA DGX B200 with 1,440 GB of GPU memory. The NVFP4 version for Blackwell GPUs is reported at about 525 GB of weights, while Yotta Labs lists 610 GB on disk. The model card itself recommends the Mistral API if local serving falls short. In short, open weight does not mean free to run: there may be no license fee, but GPUs, maintenance and expertise are on you.

What do independent measurements say about Mistral's claims?

Independent measurements paint a more cautious picture. Mistral positions Large 4 as the best open-weight coding model, citing 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3, with 47% fewer tokens per task on average. On Artificial Analysis's model page, as of October 31, 2026, its general Intelligence Index is 26, ranked 9th of 22 models. Speed on the Mistral API is 161.3 tokens per second, with 1.42 seconds to first token. That suggests strength in coding but not leadership overall. Scores such as Terminal-Bench are partly Mistral's own reporting, so they should not decide the matter alone.

What should you check before using Mistral Large 4?

Before using Mistral Large 4, settle three points: data location, language quality and latency.

Data location: A model trained on European infrastructure tells you nothing about where your API request is processed. If you handle personal data of people in the EU, Article 28 GDPR requires processing by a processor to be governed by a contract, so get the processing terms and the region in writing.
Language quality: The model card claims fluency in 160+ languages, including all official EU languages. If your customers write in a language it does not name, such as Turkish or Arabic, measure quality instead of assuming it.
Latency: Large 4 outputs text, and its measured time to first token is 1.42 seconds. On a phone call, speech-to-text and text-to-speech come on top of that. So the model alone is not a live voice AI assistant; latency has to be measured end to end.

How much does an AI assistant cost per month on Mistral Large 4?

In the example below, an AI assistant handling 10,000 conversations a month costs $58.80 on the Mistral Large 4 API without caching and $32.34 with it. We assume an average of 6,000 input and 800 output tokens per conversation; these are our example figures, not numbers published by Mistral.

Input: 10,000 × 6,000 = 60 million tokens; 60 × $0.70 = $42.
Output: 10,000 × 800 = 8 million tokens; 8 × $2.10 = $16.80.
Monthly total: $42 + $16.80 = $58.80.
If 70% of input comes from cache: (18 × $0.70) + (42 × $0.07) + $16.80 = $32.34.

This covers model usage only. Integration, testing, human handoff, WhatsApp or phone channel fees and maintenance are not included. The default reasoning mode is High, so real output can differ; update the math with your own logs. If the model bill stays at this scale, the deciding factors should be accuracy and data control rather than price.

What does this news mean for your business?

For most businesses, Mistral Large 4 means "we have more options", not "let's buy GPUs". Take a mid-sized law or accounting firm that reviews long contracts and answers repetitive client questions. The 1M-token context is attractive for long documents, and keeping client data in-house is now possible in principle, but an eight-GPU setup is hard to justify at that size. A sensible first step is comparing the model with your current one through the API, using documents that contain no personal data.

Who it suits, and who it does not:

A good fit: Companies that cannot send data outside and have or can build GPU infrastructure, software teams running coding agents, and legal or finance teams working with long document sets.
Not a good fit: Small businesses with low volume and no GPU infrastructure gain little practically from open weights. Businesses that need live phone conversations should not treat the model as sufficient on its own.

Three steps you can take this week:

Run 50 real customer questions through the Mistral API or OpenRouter alongside your current model and score the accuracy.
Pull your monthly conversation and token volume from your logs and compare it with the math above.
Get the processing country and contract terms from your provider in writing.

Choosing a model is only one part of a setup. To connect an assistant to your website, see our AI assistant installation guide. We summarized the business return in what an AI assistant brings to a business. For multi-step processes, see our AI agent service, and for a real phone deployment, our meat wholesaler case study.

Sources

Last checked: October 31, 2026.

Next step

To work out the right model and setup for your business, visit our AI assistant service page, request a short discovery call through our contact form, or call +90 850 840 08 23 and talk live with Azra, our voice assistant.

Frequently asked questions

Is Mistral Large 4 free?

Downloading the weights and using them internally requires no license fee. API usage is paid: $0.70 per million input tokens and $2.10 per million output tokens. If you run it yourself, hardware and maintenance costs are yours.

Can Mistral Large 4 run on a normal computer?

Practically, no. The model card examples target an eight-GPU node, and reported checkpoint sizes range from 525 GB to 1.06 TB. GGUF versions exist, but measure performance on your own hardware before using them in production.

Which languages does Mistral Large 4 support?

The model card claims fluency in 160+ languages, including all official EU languages. Languages it does not name should be tested with your own customer questions before going live.

Should we move our existing AI assistant to Mistral Large 4?

Only if your own tests show better accuracy, or there is a clear gain in data control or cost. Changing models means revalidating instructions, tool calls and test scenarios.

Mistral Large 4 Goes Open Weight: What Changes for Businesses, and What Doesn't