Mistral as Europe's sovereign default: Large 4, the EU API and what you can run yourself

Mistral now sells sovereignty as a whole stack, not a single model. In four months it published the weights of a trillion-parameter flagship it says it trained in its own European data centres, made EU processing a standard API option, started renting European GPU capacity and raised €3B to pay for it. No other European lab covers that many layers. The pitch has gaps, though: the flagship's licence carries a commercial clause, EU processing costs 10% more, the newest speech models are API-only, and running Large 4 yourself takes a server or a 512 GB Mac. This post maps what Mistral offers at each layer, what you can actually take home, and which Mistral model fits which job.

Our Aleph Alpha Kolibri post promised a Mistral follow-up. This is it.

Key takeaways

• Mistral Large 4 is open-weight since 29 October. It is a mixture-of-experts model with about 1 trillion parameters in total and 52B active per token, a 1M-token context and image input. Mistral says it trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres.
• The licence is permissive, with one condition. The Mistral Ecosystem Open-Weight License is a modified MIT licence. You can use, fine-tune and build products on Large 4. A company that runs a model-as-a-service business, or whose affiliate does, needs a commercial licence from Mistral before using it commercially at all; internal use is exempt.
• The API is the other half of the pitch. Mistral Studio (formerly La Plateforme) has offered EU or US regional endpoints since 11 August, billed at 1.1× list price, and it now also serves third-party open models such as GLM-5.3 and DeepSeek V4.1 Flash from the same infrastructure.
• You can run Large 4 on one machine only at 2-bit. The official GGUF files range from 298 GB to 2,099 GB; the llama.cpp maintainers say the two 2-bit files fit 512 GB of unified memory.
• Openness is selective. Voxtral 3, Mistral's newest speech family, has no public weights, and Large 4 was API-only for three weeks before the weights came out.

What Mistral shipped since July

Read the last four months in order and a pattern shows: some models went out as open weights on day one, others as API previews, and the platform and the money arrived in between.

Three of the nine events are downloads. Leanstral 1.5 (2 July) is a Lean 4 proof agent, and Shieldstral 1.0 (4 August) is Mistral's first open-weight moderation classifier; both are Apache-2.0. Mistral Large 4 followed on 29 October, three weeks after its API-only preview on 6 October. In the same period Voxtral 3 (20 October) shipped through the API only, and on 31 October Mistral deprecated the Mistral Large 3 and Mistral Medium 3.5 APIs, with retirement on 30 April 2027. The weights of both stay downloadable.

The pitch: sovereignty as a full stack

"Sovereign AI" usually means one of three things: data stays in a jurisdiction you choose, the model is one you can keep, or the supplier answers to European law. Our explainer on sovereign AI separates them. Mistral now answers all three with a product at every layer:

• Weights you can keep. Large 4, Small 4, Ministral 3, Devstral 2, Leanstral 1.5 and Shieldstral 1.0 are all downloadable. If Mistral changed its terms or disappeared, these copies would keep working.
• Training on its own European compute. Mistral says it trained Large 4 from scratch in its own European data centres.
• An API with a jurisdiction switch. Since 11 August, Mistral Studio's regional endpoints let you choose EU or US processing, and a Priority Tier adds a 99.5% uptime SLA for production traffic. Mistral says it is the only European AI lab offering both a choice of processing region and an SLA-backed service level. Regional processing is billed at 1.1× list price.
• A catalogue beyond its own models. The same endpoints serve Z.ai's GLM-5.3 and DeepSeek's V4.1 Flash-1017. That turns Mistral into a European host for the strongest Chinese open-weight models, which is a different thing from a European model, and worth keeping apart in procurement documents.
• Reserved capacity. Mistral Compute rents dedicated GPUs on EU soil with managed Slurm and Kubernetes, and Mistral plans up to 1 GW of capacity by 2030.
• Capital to keep going. The €3B Series D on 8 September valued Mistral above €21B. Mistral says it now works with more than 125 enterprises in 20 countries.

Other European projects in our library, Aleph Alpha's Kolibri, Switzerland's Apertus, EuroLLM and Poland's Bielik, give you weights. None of them sells its own pay-per-token API, its own GPU capacity and a trillion-parameter flagship (you can rent Apertus per token from third parties such as Public AI). That is the real difference, and it is why Mistral is the default name in most European procurement conversations I see at workshops. Our sovereign AI hub compares the deployment options in more detail.

Mistral Large 4 in one table

Parameter counts disagree slightly between sources: the model card says about 1T, and Mistral's API docs say 1.05T including the vision encoder. Full specs, licence badges and runtime notes are on the Mistral Large 4 model page.

On benchmarks, every number is Mistral's own. For the released weights, Mistral reports 48% on the Artificial Analysis Coding Agent Index (level with GLM-5.3 while using 47% fewer tokens per task), 41% on Terminal-Bench 4.0 (up from 28% for the preview), 19.7% on FrontierCode 2.0 and 84.5% on BrowseComp. Earlier coverage often quotes the preview's figures; don't mix the two. The model card puts Large 4 at 26 on the Artificial Analysis Intelligence Index v5, which Mistral calls the highest among western open-weight models. Leading Chinese open-weight models score higher on the same index.

The security angle is deliberate. Mistral reports 51% on the Artificial Analysis Cyber Index, a 96.6% refusal rate on malicious cyber prompts and 97.1% on the B3 agent-security benchmark, and it launched Large 4 with quotes from ENISA and a long list of security vendors. If you are evaluating local models for security work, our report on AI cyber threats and local AI covers the wider field.

Can you run Large 4 yourself?

Yes, but almost certainly not on the machine you're reading this on. Large 4 needs its full trillion parameters in memory even though only 52B work on each token: a mixture-of-experts model saves compute, not memory. Our MoE guide explains why.

There are four practical routes:

1. A 512 GB Mac, at 2-bit. The llama.cpp maintainers publish GGUF builds of Large 4 and say the two 2-bit files, IQ2_XXS at 298 GB and Q2_K at 381 GB, fit 512 GB of unified memory. Hugging Face measured Q2_K on one Mac Studio M3 Ultra at about 90 tokens/s for prompt processing and 15 tokens/s for generation. That is usable for a single person, and it is the only single-box route. It is tight: macOS lets the GPU use about 75% of unified memory by default, which the Q2_K file nearly fills, so plan on raising that limit. Unsloth also publishes 1-bit and 3-bit files in between. Two-bit quantization costs quality; our quantization guide explains the trade-off, and our Mac Studio 512 GB post covers raising the limit and the other big MoE models that fit. The Mac Studio M5 Ultra 512 GB page has the machine's specs; it is announced, not yet on sale.

2. A multi-GPU server. Mistral's own vLLM example serves the FP8 weights across eight GPUs, and the NVFP4 weights run on four NVIDIA Blackwell GPUs with SGLang. Mistral also ships a built-in speculative-decoding head it says lifts throughput up to 1.9×. Our vLLM setup guide covers the serving side.

3. Rented GPUs. If you want to test Large 4 on your own prompts before buying anything, a few hours on a rented eight-GPU node is the cheap way to find out whether it beats the model you already run.

4. The API. At $0.70 per million input tokens, the hosted model is cheaper than any of the above for most teams, and the EU endpoint keeps processing in Europe for 10% more. The model card itself recommends the API if your local serving falls short.

Check the exact fit for your hardware in the analyzer; it uses the same memory model as the rest of the site. A context window is not a memory figure: the KV cache for long prompts comes on top of the file sizes above, and the VRAM guide shows how to budget it.

If Large 4 is too big, the rest of the family is not. Mistral Small 4 is a workstation-class MoE, Ministral 3 runs on laptops and phones, and Devstral 2 is the coding specialist.

The API side: EU processing, third-party models

Downloads are free; Mistral earns its money from the API and enterprise contracts, so the API is where the sovereignty claim gets tested every day. These are the list prices on 3 November 2026, per million tokens, before the 1.1× regional multiplier:

Two things stand out. First, Large 4 costs less per token than GLM-5.3 on Mistral's own price list, though keeping processing in the EU adds 10% to either. Second, hosting Chinese open-weight models on European infrastructure is a pragmatic move: it gives customers the strongest open models under an EU contract and EU processing. It also shows that, for Mistral, "sovereign" describes where and under whose contract inference runs more than who trained the model. Our cloud directory lists Mistral Studio and Mistral Compute next to the other hosted options.

Where the sovereignty story is thinner

Mistral's position is strong, and it isn't complete. If you are writing a procurement case, these are the points a sceptical reviewer will raise.

• The flagship licence is not Apache-2.0. Leanstral 1.5 and Shieldstral 1.0 are Apache-2.0; Large 4 is under a modified MIT licence with a model-as-a-service condition. If your company does not run a model-as-a-service business, the condition doesn't apply. If it does, or an affiliate does, it needs a commercial licence before any commercial use of Large 4 beyond internal use. Read the licence file, not a summary; our licence guide explains what to look for.
• The region switch covers inference, not the account. Mistral's own docs say regional endpoints do not make the control plane regional: API keys, billing and usage analytics may be handled outside the chosen region, and the default global endpoint commits to no location at all.
• Openness is selective. Voxtral 3 is API-only, and running it on your own infrastructure goes through Mistral's team. Mistral OCR and Codestral are closed API models. The earlier Voxtral TTS weights are non-commercial, as our voice hub notes. Large 4 itself was API-only for three weeks.
• API models retire. Large 3 and Medium 3.5 got six months between deprecation and retirement. That is a normal API lifecycle, and it is exactly why the downloadable weights matter: they are your exit if a hosted model you depend on goes away.
• The hardware is American. Large 4 was trained on NVIDIA GPUs, the NVFP4 weights target NVIDIA Blackwell, and Microsoft Foundry hosts the model for Azure customers. European data centres running American chips are still a dependency, just a different one from American data centres.
• The money is international. Samsung led the Series D, with EQT's Scaleup Europe Fund and PSG Equity as co-leads; US investors such as BlackRock and Advent joined, as did the Grand Duchy of Luxembourg.
• Capability still trails the best open models. By Mistral's own reading of the Artificial Analysis index, Large 4 leads western open-weight models but not the leading Chinese ones. And every benchmark in the launch material is vendor-reported.

None of these makes Mistral a weak choice. They mean "sovereign" should be a checklist you score per layer, not a label you accept whole.

Which Mistral model for which job

Leanstral's hosted API was retired on 30 September, so the open weights are now the only way to use it. Shieldstral takes your policy as a yes/no question in each request and returns a safety score, which makes it a good companion to any model in this table.

What this means for European teams

The pattern I recommend in workshops is "API with an exit": start on Mistral's EU endpoint, keep prompts and evaluation sets portable, and make sure the model you depend on has downloadable weights you have actually tested on your own hardware, even once. The EU endpoint covers where inference runs on day one; the weights cover supplier risk. With Large 4, both halves exist for Mistral's newest flagship from its first month.

Before you commit:

• check data processing and transfers against our GDPR comparison;
• check your obligations as a deployer in our EU AI Act overview;
• price self-hosting against the API with the on-premise TCO guide;
• and read the bigger picture in our report on Europe's compute gap.

Frequently asked questions

Methodology and sources

Researched 3 November 2026. We read Mistral's announcements, API documentation and price list, the model cards and licence files on Hugging Face, the llama.cpp maintainers' and Unsloth's GGUF repositories and Hugging Face's launch blog. We did not run Large 4. Benchmarks are vendor-reported, the only speed figure is Hugging Face's measurement, and file sizes are the publishers'. Where sources disagreed, total parameter count and preview versus final benchmarks, we used the final model card and say so in the text.

• Mistral: Le Chonk: Built for Coders, Trusted by Defenders (29 Oct 2026), read 2026-11-03
• Mistral: Introducing Mistral Large 4 (preview, 6 Oct 2026), read 2026-11-03
• Mistral Large 4 model card (FP8), read 2026-11-03
• Mistral Ecosystem Open-Weight License (LICENSE file), read 2026-11-03
• Mistral Large 4 technical report (added 2 Nov 2026), read 2026-11-03
• Mistral Large 4 NVFP4 model card, read 2026-11-03
• Mistral docs: Mistral Large 4, read 2026-11-03
• ggml-org GGUF repository and card (llama.cpp maintainers), read 2026-11-03
• Unsloth GGUF repository (file tree), read 2026-11-03
• Hugging Face blog: Mistral Large 4, read 2026-11-03
• Mistral docs: models and deprecation table, read 2026-11-03
• Mistral docs: API pricing, read 2026-11-03
• Mistral: In-region inference, open models, and new European infrastructure (11 Aug 2026), read 2026-11-03
• Mistral docs: Regional inference, read 2026-11-03
• Mistral docs: Priority Tier, read 2026-11-03
• Mistral raises EUR 3B (8 Sep 2026), read 2026-11-03
• Mistral: Voxtral 3 (20 Oct 2026), read 2026-11-03
• Mistral Compute product page, read 2026-11-03
• Mistral: Leanstral 1.5 (2 Jul 2026), read 2026-11-03
• Leanstral 1.5 model card, read 2026-11-03
• Mistral: Introducing Shieldstral (4 Aug 2026), read 2026-11-03
• Shieldstral 1.0 3B model card and config, read 2026-11-03
• Public AI gateway model list (serves Apertus per token), read 2026-11-03