On October 6, 2026, Mistral AI released Mistral Large 4 as a public preview—a language model with over a trillion parameters, whose weights are scheduled to be available for download by the end of October. For IT managers and CIOs in industry and retail with their own AI plans—such as for reviewing technical drawings in plant engineering or evaluating customer complaints at an automotive supplier—this raises a concrete budget question: Is it worth budgeting for hardware to run this model, or will the API or a smaller model suffice?
Mistral Large 4 is a data center model when you run it in-house: even in 8-bit, the weights take up about 1 TB of GPU memory, which corresponds to a server with eight B200s. Do not order hardware for it yet. The license, delivery formats and independent tests of the final weights are still missing (as of October 8, 2026). It makes sense to test the preview with non-critical data, decide once the license is published, and check in parallel whether a smaller model on one or two cards already handles the task.
| Characteristic | Mistral Large 4 |
|---|---|
| Publisher | Mistral AI, Paris |
| Architecture | Mixture of Experts, text and images as input, text as output |
| Parameters | 1.05 trillion total, 49 billion active per token (52 billion including embedding and output layers) |
| Context Window | According to Mistral, 1 million tokens; Artificial Analysis cites 512,000 tokens for the preview |
| Status | Research Public Preview via API since October 6, 2026 |
| Weights | announced for release by the end of October 2026; Hugging Face lists October 31, 2026 |
| License | Not yet published |
| API List Price | $1.36 per 1 million input tokens, $4.18 per 1 million output tokens; half price during the first two weeks |
What Mistral Announced
Mistral describes Large 4 as its largest model to date and, according to the company, has trained it on 3,800 graphics processing units in its own data centers in Europe (Mistral AI, Oct. 6, 2026). The manufacturer lists programming, cybersecurity, finance, and law as areas of application, as well as image-processing tasks in engineering and manufacturing—such as inspecting mechanical components in technical drawings. The performance figures on the announcement page are provided by Mistral itself.
The first independent measurement is now available: Artificial Analysis rates the preview at 38 points on its own Intelligence Index, just behind DeepSeek V4.1 Flash (39). According to this assessment, a test task costs $1.13 via the API—more than four times as much as comparably powerful open-source models such as GLM-5.3-Flash, which costs $0.25 (Artificial Analysis, Oct. 6, 2026). The licensing status is still unclear: Mistral Large 3 is licensed under Apache 2.0, while VentureBeat reports that Large 4 will use a proprietary Mistral license, the terms of which are not yet known (VentureBeat, 06.10.2026).
The graphics memory must hold all of the parameters, even though only 49 billion of them are processed per token.
Mistral AI, Documentation, and Hugging Face, October 2026
What hardware in-house operation requires
In mixture-of-experts models, memory determines the hardware, not computing power. The table shows the memory requirements for the weights alone; context memory for concurrent queries must be added to this.
| Format | Weights | Compatible System |
|---|---|---|
| BF16 (16-bit) | about 2.1 TB | more than one server with eight cards, such as two DGX B200s |
| FP8 (8-bit) | about 1.05 TB | a DGX B200 with about 390 GB of spare capacity; eight H200s (1,128 GB) with only about 80 GB of spare capacity |
| 4 bits | about 0.53 TB | eight H200s with headroom; eight RTX PRO 6000s (768 GB) possible on paper |
It remains to be seen what formats Mistral will support. For Mistral Small 4, the manufacturer has released its own 4-bit format (NVFP4). Operation also comes at a cost: NVIDIA specifies a maximum power consumption of approximately 14.3 kW for a DGX B200. Under continuous full load, that would amount to 125,268 kWh per year; at 17.2 cents per kWh—the average price for small to medium-sized industrial companies in 2026 according to the BDEW—that would be approximately 21,500 euros in electricity costs, not including cooling. That is an upper limit.
Calculation Example: API or Your Own Server
A plant engineering company wants to use an assistant that searches manuals and trouble reports to respond to service requests from its technicians. The company expects 1,500 requests per workday, each with 8,000 tokens of context and a 700-token response.
- Input: 12 million tokens per day at $1.36 per million equals $16.32.
- Output: 1.05 million tokens per day at $4.18 per million equals $4.39.
- A total of $20.71 per day, which amounts to about $4,560 per year at the list price, based on 220 working days.
The power consumption limit of the server alone is several times higher, and then there’s the purchase price: A dealer quotes a starting price of 149,000 euros net for a system with eight RTX PRO 6000s (dealer quote, 08/2026). From an economic standpoint, operating your own Large 4 system isn’t cost-effective at this volume. It’s only worthwhile if the data must remain on-premises and no smaller model can handle the task, or if the volume is many times greater—for example, when a corporation shares a single server across all its locations.
| Option | Effort | Costs | Risk |
|---|---|---|---|
| Test the preview via the API | a few days | Depends on usage; see above for list prices | Data is sent to Mistral, preview status; only non-sensitive or anonymized test data |
| Do nothing: Wait for the weights, license, and tests | none | none | minimal; if purchased later, delivery times of 6 to 12 weeks will delay the start |
| Operate Large 4 Yourself | Sizing, data center with approximately 14 kW per server | six-figure investment plus electricity | License and formats are still unknown |
| Operate a smaller, open model yourself | 2 to 4 weeks of sizing and testing | one or two cards instead of eight | may perform less well with long agent processes and graphics-intensive tasks |
What Mistral Large 4 is suited for
Large 4 is particularly useful where smaller models reach their limits: multi-stage agent workflows spanning multiple systems; tasks that combine images and text—such as drawings, inspection reports with photos, or damage reports from the field; and security analyses. In our assessment, a model that runs on one or two cards is sufficient in many cases for searching manuals, standards, and contracts. Kolibri from Aleph Alpha, for example, is licensed under Apache 2.0 and, according to the manufacturer, runs on a single H200.
Test Large 4 against a smaller modelWe will run 50 tasks from your company—using approved or anonymized data—on the Mistral Large 4 preview and on a model that runs on one or two cards. Within ten business days, you will receive a comparison of the results, the hardware requirements for both options, and a cost estimate based on your request volume.
Frequently Asked Questions
When will the weights of Mistral Large 4 be released?
Mistral has announced that the weights will be released by the end of October 2026; the Hugging Face page lists October 31. VentureBeat reports the date as October 27.
Is it permitted to operate Mistral Large 4 commercially on one's own?
That is still open, because Mistral has not published the license. Its predecessor Mistral Large 3 is licensed under Apache 2.0; for Large 4, VentureBeat reports a custom Mistral license. Check the license text before you order hardware for in-house operation.
Does Mistral Large 4 run on a workstation or a DGX Spark?
No. Even in 4-bit mode, the weights take up about 530 GB; a DGX Spark has a maximum of 128 GB, and a workstation card has 24 to 96 GB. To run it yourself, you need a server with eight data center cards.
The Next Step
torck designs and operates open language models in the data centers of industrial and commercial clients, handling everything from selection to operation. During the initial consultation, we determine whether your use case requires a model of the “Large 4” size or if a smaller model will suffice. AI Consulting with On-Premises LLMs.
Read more: When buying hardware pays off is covered in our article on AI Hardware 2027. The hardware the German model requires is described in the article on Kolibri by Aleph Alpha. Smaller models from Mistral, Llama and Qwen are compared in our Open-Source LLM Comparison.
German 
English 
French