Mistral Large 4 and the Real Cost of Running AI In-House

What Mistral shipped on October 6, and what is still missing

On October 6, 2026, Mistral AI opened a paid public preview of Large 4, a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49 billion active parameters, according to AI News. The preview runs through Mistral Studio. Mistral says the weights will follow by the end of October, and The Next Web reports a date of October 27, after which customers can run the model on their own servers or use Mistral's API in a region they choose, including Europe.

What is missing matters as much as what was announced. At launch there were no downloadable checkpoints, no licence file and no repository, as OrcaRouter pointed out, and Implicator noted that Mistral's Hugging Face page listed no Large 4 checkpoint. Mistral itself describes its benchmark results as preliminary and says reinforcement learning is still ongoing. In practice, Large 4 is today an API product with a promise attached.

The preview pricing gives you a baseline for any comparison: $1.36 per million input tokens and $4.18 per million output tokens, currently discounted to $0.68 and $2.09, with cached input at $0.14 (discounted to $0.07). Keep those numbers in front of you. Every self-hosting business case has to beat them, not just feel safer than them.

How good is it when someone other than Mistral measures it

Mistral's own launch post claims 93% of Cybench challenges, 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and a Coding Agent Index score of 49.8%, which it says puts Large 4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. It also says the model resists 93.3% of attacks on Lakera's B3 AI Security Benchmark. These are vendor numbers on vendor-selected tests, and they deserve the usual discount.

Independent rankings are more sober. Large 4 scored 38 on the Artificial Analysis Intelligence Index, the highest of any open model built outside China, but seven Chinese open models score higher, and Artificial Analysis lists the preview as proprietary until weights appear. On the Vals Index it ranks #32 of 44 models at 48.05%, while Google's closed Gemini 4 Argon, released September 30, sits at #1 with 68.90%. Large 4 trails the closed frontier by about 21 points on the Vals Index. For buyers who want a European vendor and prefer not to deploy Chinese weights, it may still become the strongest option once the weights ship.

The spread across tasks is the more useful signal. Its best independent results are on legal-agent and medical-scribing tasks: #6 of 75 on Harvey's Legal Agent Benchmark and 80.43% on MedScribe. If your workload resembles one of those, the overall rank may understate what you get. If it does not, those two results tell you little.

Leaderboards are dominated by English tasks; if your documents are in Turkish or another language, build your own test set. Mistral says it trained Large 4 on more than 160 languages, including every official EU language, but that is a training claim, not a measured result for your summarisation or extraction work.

What open weights actually let you do

Open weights mean you can download the model file and run it wherever you like, subject to the licence. VentureBeat reports that Mistral is pitching Large 4 as a foundation that enterprises and governments can run and customise themselves, including on sovereign infrastructure. In concrete terms, that opens up a few things an API cannot give you.

It is worth noticing that Mistral also offers regional API endpoints and zero-data-retention options. For many companies that middle path covers the real requirement, which is usually 'our data must not be stored or used for training abroad' rather than 'our data must never leave the building'. Be precise about which of those you actually need before buying hardware.

• On-premise deployment: prompts, documents and outputs never cross your network boundary.
• Data residency: you choose the country and the data centre, not the provider's region list.
• Fine-tuning: you can adapt the model to your terminology, formats and languages, if the licence allows it.
• Version control: hosted models can be updated or retired on the vendor's schedule, while a self-hosted model stays at the version you pinned until you decide to change it.
• Cost predictability: at high, steady volume you pay for hardware and staff instead of per token.

The licence will decide more than the benchmarks

Mistral Large 3 shipped under Apache 2.0, one of the most permissive licences there is. Large 4's weights are expected under a custom Mistral licence instead, and the text has not been published. That single document will determine whether most of the benefits above are real for your company.

As Beri.net explains, custom open-weight licences are typically where revenue thresholds, acceptable-use lists and limits on serving the model to third parties appear. Until the text is out, nobody can say whether a bank may fine-tune Large 4 or whether a systems integrator may host it for clients. 'Open weights' is not the same as open source, and the difference is often commercial.

One development team has already put its Large 4 self-hosting plans on hold in a public GitHub issue. Its conditions for re-opening the work include reading the licence for field-of-use, revenue or monthly-active-user caps and understanding the EU AI Act duties for general-purpose AI models. That is a sensible sequence. Do not order GPUs before your legal team has read the licence.

The costs that do not show up on the model card

Hardware comes first. Large 4 is too large to run on a desktop or laptop, and the GitHub team above estimates it needs multi-GPU H100-class nodes. The mixture-of-experts design means only 49 billion parameters are active for each token, which helps speed, but the full set of roughly a trillion parameters still has to sit in GPU memory. Quantized versions will shrink that footprint, usually at some cost in quality, so test the quantized model on your own tasks rather than assuming it behaves like the full one.

Then compare against the API. As a rough baseline, 100 million input tokens plus 20 million output tokens at list price cost $136 + $83.60 = $219.60, and about half that at the current discounted preview prices. Summarisation and drafting shift the mix toward output tokens, so model your own ratio rather than borrowing this one. A small or mid-sized company processing internal documents may never spend enough on tokens to justify a dedicated GPU node, let alone the people to run it.

Self-hosting tends to pay off only when volume is high and steady enough to keep expensive hardware busy, or when data rules leave no alternative. The less visible costs are often bigger than the hardware. Mistral's 93.3% figure on Lakera's benchmark describes the model, not your deployment: once you self-host, the serving stack, access control, logging and patching are your attack surface and your responsibility.

• Operations staff who can run inference servers, monitor them and be on call.
• Evaluation work every time you upgrade, quantize or fine-tune.
• Security hardening, access logging and incident response for the serving stack.
• Idle capacity: GPUs cost the same whether they are busy or not.
• Legal review of the licence and any obligations under the EU AI Act.

Governance gaps that self-hosting will not close

The 2026 National Cybersecurity Alliance survey, run with CISA across 1,000 small business leaders in 10 industries, found that 87% use AI tools but only 46% have formal guidelines for that use. MainNerve also warns that consumer-tier AI tools often give the provider broad rights to use inputs for model training. That is the more likely data leak in most companies: an employee pasting a contract into a free chatbot.

A self-hosted model does nothing about that on its own. It controls where data goes once it reaches your model, not whether staff use your model at all. If you have no usage rules today, write them before you build infrastructure, because the infrastructure will not enforce itself.

Open weights let you pin and audit a specific model version, whereas hosted vendors disclose only what they choose, and you take on the testing duty yourself. This week showed the limits of disclosure: at the October 5 NYC Council hearing with sworn testimony from Anthropic, OpenAI, Google and Meta, OpenAI said it is reviewing potential past AI agent misalignment incidents back to November 2025, and the council left with many unanswered questions.

Meeting transcripts, HR records and customs documents

Meeting recordings are a strong case for local processing. Board discussions, HR conversations and negotiations contain strategy and personal data in the same file. Recordings of identifiable people are generally treated as personal data under GDPR and KVKK; confirm with counsel. A common pattern is to run speech-to-text and summarisation on your own servers, so neither audio nor transcript leaves the network, and here language quality matters more than raw model size, because a summary is only as good as a transcript that got the names, numbers and accents right.

HR data is the case where legal review usually decides. Employee records, leave history and payroll may include sensitive categories such as health information, and cross-border transfer rules can make cloud processing a lengthy approval exercise. Most HR tasks, such as extracting fields from forms or drafting standard letters, do not need a trillion-parameter model. A smaller open model on a single server is often enough, and much cheaper to run.

Customs documents sit in between. Invoices, packing lists and declarations are commercially confidential but often contain little personal data, and volumes tend to be uneven. A regional API with zero data retention may satisfy the requirement without any GPUs of your own. Large 4's legal-agent result does not tell you how it handles trade paperwork, so test extraction on a sample of your own documents before committing either way.

A checklist for deciding between self-hosting and an API

Run through these questions in order. If you answer no to the first question and yes to the second, start with a regional API with zero data retention. If you answer yes to the first, or no to the second, the remaining questions tell you whether you are ready to self-host or should look for a vendor that deploys on your premises.

My view for the rest of 2026: evaluate Large 4 through the preview API on synthetic or non-sensitive data, wait for the licence text, and then decide. Self-hosting is worth it when data control is non-negotiable and the workload is steady. It is not worth it as a gesture, or as a substitute for usage rules you have not written yet.

• Does law or contract forbid this data from leaving your infrastructure?
• Would your legal counsel accept a regional API with zero data retention?
• Have you read the licence for field-of-use limits, revenue or user caps and restrictions on serving third parties?
• Is your volume high and steady enough to keep GPUs busy most of the day?
• Do you have people who can run, secure and upgrade an inference stack around the clock?
• Have you tested quality on your own documents, in your own languages?
• Do you need the largest model, or would a smaller open model handle the task?
• Do staff already have written rules for which AI tools they may use and with what data?