"Can I paste the Polish supplier contract into the chatbot? I can't read clause 7."
"Which chatbot?"
"The one everyone uses."
"Then no. Not with the prices in it, and not until legal knows where it goes."
I have heard some version of that exchange in every company I know. The buyer is not wrong to ask: reading terms in five languages is slow work. The manager is not wrong to refuse: a supplier contract is exactly the kind of text that should not travel without a plan.
On 6 October Mistral, the Paris AI company, opened a preview of Mistral Large 4. It is not the cleverest model on the market. What makes it interesting for that buyer is where it runs, and where it will soon be able to run.
What Mistral Large 4 is
Mistral Large 4 is Mistral's new flagship model, released as a public API preview on 6 October 2026. Mistral says it is "natively multimodal", reads more than 160 languages including every official EU language, and was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres.
Under the bonnet it is a mixture-of-experts model. Think of a large firm of specialists where only the few who know the subject work on each question. The model has about 1 trillion parameters in total, but only around 50 billion are active for any one word it writes. That keeps it fast and cheap for its size.
What you can use today, and what is coming:
- API preview in Mistral Studio, run in Mistral's European data centres under European law. MindStudio reports it is not yet in Le Chat, Mistral's chat app.
- Price: $1.36 per million input tokens and $4.18 per million output tokens, with half those rates during the preview, according to Beam AI's reading of the model card.
- Open weights promised by the end of October, after more red-team testing. Reuters reported 27 October, via Tech Journal.
- Tool calling and strict JSON output, so it can feed a spreadsheet or another system without reformatting.
- A one-million-token context window on paper. Beam AI found the preview API currently shows about 524,000.
The benchmark figures are Mistral's own, so treat them as self-reported. It claims 59.9% on AutomationBench (office automation tasks), 61.7% on DeepSWE v1.1 (coding) and strong results on cybersecurity tests. Independent scoring is cooler. Artificial Analysis gives it 38 on its Intelligence Index, behind the Chinese open-weight models GLM-5.3 (45) and Kimi K3 (44), as reported by Beam AI.
So the case for Large 4 is not raw intelligence. It is the combination: close to frontier size, European operation now, and weights you can download soon.
What it means for supply chain
Most AI discussions in supply chain start with "which model is smartest?" For contracts, quality files and customs papers, the first question is a different one: where does the text go?
| Where the model runs | Who handles your data | Fits for |
|---|---|---|
| A US-hosted chatbot or API | The provider, under its terms and US law | Public or anonymised text |
| An EU-hosted API (Large 4 preview today) | Mistral, in its European data centres under European law | Supplier documents your policy allows to leave the company under a contract |
| Your own servers (Large 4 weights, from end October) | Only your company | Prices, contracts and anything legal will not let out |
Large 4 is the first model of this size that may sit in all three rows with the same behaviour. That gives a team one model to test, from pilot to production.
Where I would use it:
- Supplier terms in many languages. Pull Incoterms, payment terms, lead times and price-revision clauses from contracts in Polish, Italian or Dutch into one English table.
- Quality and compliance files. Read supplier certificates and declarations as PDFs, including scanned ones, and list what is missing or expired.
- Engineering drawings for spare parts. Mistral says the model is good at "engineering diagram inspection". Worth testing on part drawings when cleaning up the item master.
- Trade and customs papers kept in Europe. Commercial invoices and origin statements can be checked without leaving European infrastructure.
- A private base model. Once the weights are out, an IT team can run it inside the company network for jobs where no outside API is acceptable.
The use I would test first is the buyer's problem from the opening. Here is a prompt that does exactly that.
Try this: one standard, five languages
Run it in Mistral Studio with Large 4, which is where the preview is available. Then run the same prompt in the model you use today and compare. The clauses are illustrative and belong to a fictional bicycle maker; never test a preview model with real contracts.
You are a procurement analyst at a fictional bicycle maker.
Task: check five supplier contract clauses against our standard terms. The clauses are in five languages. Work in English.
Our standard terms:
1. Incoterms 2020 DAP, buyer's plant.
2. Payment 60 days end of month.
3. Lead time at most 10 working days.
4. Price revision at most once a year, with at least 3 months' notice.
Clauses:
A (German, frames): "Lieferung DAP Werk des Käufers (Incoterms 2020). Zahlung 60 Tage nach Monatsende. Lieferzeit 8 Arbeitstage. Preisanpassung einmal jährlich mit 3 Monaten Vorankündigung."
B (Italian, saddles): "Resa EXW stabilimento del fornitore (Incoterms 2020). Pagamento a 30 giorni data fattura. Termine di consegna: 10 giorni lavorativi. Prezzi adeguabili ogni trimestre in base al costo delle materie prime."
C (Polish, wheels): "Dostawa DAP zakład kupującego (Incoterms 2020). Płatność 60 dni od końca miesiąca. Czas dostawy 15 dni roboczych. Ceny stałe przez 12 miesięcy, zmiana z 3-miesięcznym wyprzedzeniem."
D (Dutch, packaging): "Levering DAP fabriek koper (Incoterms 2010). Betaling 45 dagen na factuurdatum. Levertijd 5 werkdagen. Prijswijziging met 1 maand aankondiging."
E (French, batteries): "Livraison DAP usine de l'acheteur (Incoterms 2020). Paiement à 60 jours fin de mois. Délai de livraison : 10 jours ouvrés, sauf pour les cellules de batterie : 12 semaines. Révision des prix une fois par an avec un préavis de trois mois."
Constraints: quote the original words for every deviation. Do not guess what a clause means if it is ambiguous; flag it.
Output:
1. A table: supplier, term, our standard, their clause (English translation), compliant yes/no.
2. The same result as JSON, one object per supplier, with a list of deviations.
3. A short list of the assumptions you made.What you should see
A good answer finds eight deviations, and supplier A is clean.
| Supplier | Deviations |
|---|---|
| A, German | None |
| B, Italian | EXW instead of DAP; 30 days from invoice date; price changes every quarter |
| C, Polish | 15 working days lead time |
| D, Dutch | Incoterms 2010 instead of 2020; 45 days from invoice date; only 1 month's notice and no limit on how often |
| E, French | 12 weeks for battery cells, hidden inside an otherwise compliant clause |
The two traps are D and E. "Incoterms 2010" looks right at a glance, and the battery-cell exception in E sits after "sauf", French for "except". Those are the lines a tired reader skips at six in the evening.
Why this shows the model's power
Any strong model can translate Dutch. The test is whether it keeps five languages straight, quotes the original wording, spots a version number and an exception, and returns clean JSON in the same answer. A hasty reading marks E compliant because the 10 days comes first.
If your current model gets the same score, fine. Then the reason to choose Large 4 is not the answer. It is where the contract was read.
What it can't do, and the traps
- It is a preview. Beam AI reports reinforcement learning was still running at launch. MindStudio reports runaway reasoning that can fill the context without a clear error. Do not build a process on it yet.
- The licence is unknown. Mistral Large 3 used Apache 2.0, but Beam AI reports Large 4 may get a custom Mistral licence. Read it before you plan to self-host.
- Self-hosting is not cheap. All 1 trillion parameters must sit in memory, even though only about 50 billion work at a time. That means several high-end GPUs, not a server under a desk.
- It is not the smartest model. On independent scores it trails GLM-5.3 and Kimi K3. For hard reasoning, a closed frontier model may still do better.
- Test it on your own documents. Take ten anonymised contracts where you already know the deviations, and count what each model finds and misses.
Ask where the model runs
This week, find out which AI tools your company allows for supplier documents, and in which region they run. Then run the prompt above in two models and compare the deviation tables side by side.
For years, "where does our data go?" was the question that stopped AI projects in procurement. With a model like this, it could become one of the selection criteria.
Which documents in your work would you only hand to an AI that runs inside your own walls?
