TL;DR: On October 6, 2026, Mistral AI launched Mistral Large 4 in public preview, a 1.05-trillion-parameter model trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers, with open weights promised by the end of October 2026.
Until now, the choice often came down to this: a top-tier American model, or a European compromise. Mistral's announcement changes the terms of that trade-off, and it comes with verifiable numbers rather than a manifesto.
Key takeaways:
- Mistral Large 4, launched in preview on October 6, 2026, has 1.05 trillion total parameters with 49 billion active per token, and a 1-million-token context window.
- The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in data centers Mistral owns in Europe, and is served from that same infrastructure under European law.
- Preview pricing is $0.68 per million input tokens and $2.09 per million output tokens, half the list price; an EU regional endpoint costs 10% above list.
- Model weights are promised by the end of October 2026; Mistral claims 93.3% resistance to indirect prompt injection on the B3 benchmark.
What Mistral shipped on October 6
Mistral AI released the public preview of Mistral Large 4 on October 6, 2026, available through the Mistral Studio API, per the official announcement. The architecture is a natively multimodal Mixture of Experts (MoE): 1.05 trillion total parameters, 49 billion activated per token, a 1.6-billion-parameter vision encoder and a 1-million-token context window. The model covers more than 160 languages, including every official language of the European Union.
On benchmarks, Mistral claims the top spot among open-weight models outside China in aggregate, with measured strengths: 61.7% on DeepSWE v1.1 (software engineering), 59.9% on AutomationBench, a suite of 657 business workflows, 93% on Cybench (cybersecurity) and 42% on Dense 200 (visual grounding), one point ahead of OpenAI's GPT-6 Astra, according to Connic's breakdown. Independent evaluator vals.ai reportedly places the model ahead of GPT-6 Astra on legal and financial suites, reports Bushletter.
On pricing, the preview costs $0.68 per million input tokens, $0.07 for cached input and $2.09 per million output tokens; list pricing, which will apply at general availability, is $1.36 for input and $4.18 for output. Batch processing gets a 50% discount. Finally, Mistral has promised to publish the model weights before the end of October 2026; the license is not yet announced, whereas Mistral Large 3 shipped under Apache 2.0.
Why does an all-European training run matter?
Because sovereignty stops being a marketing claim and becomes a verifiable property of the whole chain. Mistral Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs installed in data centers Mistral owns in Europe, and the preview is served from that same infrastructure. Mistral states that its European deployment is operated "independently of other digital service providers and under European law". Until now, even European models were mostly trained on American hyperscaler clouds; where a model company was headquartered said nothing about where its model was built or run.
The second building block is open weights. A model whose weights are published can be audited, fine-tuned and hosted by any player, including a French or German provider subject only to European law. That is the difference between a contractual promise of confidentiality and a material ability to control. The caveat stands until the weights actually ship: as of October 7, 2026, only Mistral-hosted access exists, and the weight license remains unknown.
The third element is less visible but decisive for business use: resistance to attacks. Mistral reports 93.3% resistance to indirect prompt injection on the B3 benchmark and says no competitor currently scores higher, plus a global top 5 on the Artificial Analysis Cyber Index. The model is undergoing red-teaming with cybersecurity firms and state authorities. For an AI agent plugged into a mailbox or a CRM, prompt-injection resistance is not a benchmark footnote; it is the difference between an assistant and an open door.
What does it change for a European SMB?
Three things, concretely.
First, the performance-versus-sovereignty trade-off is weakening. A company handling customer data, health records or legal documents could legitimately hesitate to send them to a model operated from the United States; the price of that caution was a less capable model. With a frontier-class model trained and served in Europe, credited with strong results precisely on legal and financial suites, that hesitation has less and less technical ground.
Second, the access options are now clear, with very different costs. The table below sums up the four paths available to a business.
| Option | Cost | Sovereignty | Best for |
|---|---|---|---|
| Mistral API, preview | $0.68 in, $2.09 out, per million tokens | Mistral infrastructure in Europe | Fast testing, prototyping |
| EU regional endpoint | List price plus 10% | Processing guaranteed inside the EU, under European law | Customer or regulated data |
| Self-hosted weights (late October) | About 1 TB of GPU memory at 8-bit, plus operations | Maximum, everything stays in-house | Mid-caps and regulated sectors with an infra team |
| Through a provider or a business tool | Included in the tool's cost, on quote | Depends on the chosen host | SMBs without a dedicated tech team |
Third, judgment still applies. Self-hosting a 1.05-trillion-parameter model takes roughly 1 TB of GPU memory at 8-bit precision, more than a standard eight-H100 server provides; for a 10-to-100-employee company, that path almost never makes sense. Likewise, if your use cases are drafting and summarizing with no sensitive data involved, a smaller, cheaper model is plenty; sovereignty has a price, and it should answer a real need, not a reflex. Most of the benchmarks cited are also the vendor's own, to be confirmed by independent evaluations once the weights are out.
How this connects to my own work
This announcement lines up with a conviction I have been applying for months: with sensitive data, the AI model must adapt to the sovereignty constraint, not the other way around. On the Emma CRM for France's 3018 hotline, the national service against cyberbullying, the integrated AI is sovereign and the platform is hosted in France, precisely because conversations with teenagers in distress cannot transit through infrastructure subject to foreign law; I told that story in the 3018 case study. With Mistral Large 4, that kind of architecture gains access to a performance class that used to be reserved for American models. If you are wondering what a sovereign AI agent could automate in your business without your data leaving the EU, that is exactly the kind of scoping a 30-minute call is for, and my AI agent developer page details the approach.
The question is no longer whether Europe can produce a frontier model; Mistral just answered it with 3,800 GPUs and a European power bill. The question still open is the one that concerns you: the day your business tools embed AI deeply, will you be able to say where your data is processed, and under which law?
Frequently asked questions
Is Mistral Large 4 available today?
Yes, as a public preview through the Mistral Studio API since October 6, 2026. The model weights, which will allow self-hosting and auditing, are promised by the end of October 2026. The license for those weights has not been announced yet.
How much does Mistral Large 4 cost?
During the preview, Mistral charges $0.68 per million input tokens and $2.09 per million output tokens, half the list price ($1.36 and $4.18). The EU regional endpoint, where data is processed inside the European Union, costs 10% above list price.
Can a small business self-host Mistral Large 4?
Technically yes once the weights are released, but it requires roughly 1 TB of GPU memory at 8-bit precision, more than a standard eight-H100 server provides. For most SMBs, the realistic path is Mistral's API, the EU endpoint, or a service provider operating the model for them.