Mistral Large 4 and European AI models: licences, EU rules and GPUs

Part of our AI models guide: 11 posts on choosing and running AI models.
Mistral Large 4, the flagship of Europe's leading model maker, Mistral AI in Paris, has had downloadable weights since 29 October 2026. In its standard 8-bit (FP8) version it needs one full 8-GPU NVIDIA B200 or B300 server. Check the licence before the hardware: Small 4 and Ministral 3 are Apache 2.0, Large 4 needs Mistral's licence if your company sells model access as a service, and Medium 3.5 shuts out companies above $20 million in monthly revenue. Research groups add open models for European languages, small enough to run several copies on one server, such as EuroLLM and Switzerland's Apertus, which its makers describe as fully open. Running open weights on dedicated GPUs in Europe or the UK keeps prompts and customer data in the region you choose.
Europe's models do not top the leaderboards. They win on control: which model you may legally run, on what hardware, and where. For the US and Chinese leaders, see our guide to how many GPUs you need to run an LLM.
Europe's AI models at a glance
Parameters are the numbers a model learns (B = billion, T = trillion). Open weights means you can download them and run the model yourself (see open-weight vs closed models). Context is how much text a model can read at once, counted in tokens (the smallest units of text an AI model reads and writes, often a short word or part of one).
Model | Maker | Size | Licence | Released | Worth knowing |
|---|---|---|---|---|---|
Mistral Large 4 | Mistral AI, France | About 1T, 49B active (52B with embeddings); mixture of experts (MoE) | Mistral Ecosystem Open-Weight | Weights 29 Oct 2026 | 1M-token context; image input; 8-bit (FP8) and 4-bit (NVFP4) versions |
Mistral Medium 3.5 | Mistral AI | 128B, dense | Modified MIT, $20M monthly revenue cap | 28 Apr 2026 | 256k context; leaves Mistral's API on 30 Apr 2027 |
Mistral Small 4 | Mistral AI | 119B, 6.5B active (MoE) | Apache 2.0 | 16 Mar 2026 | Chat, reasoning, images and code in one model |
Ministral 3 (3B, 8B, 14B) | Mistral AI | 3B to 14B, dense | Apache 2.0 | 2 Dec 2025 | Mistral says the 8B fits in 12 GB at FP8 |
Mistral Large 3 | Mistral AI | 675B, 41B active (MoE) | Apache 2.0 | 2 Dec 2025 | Leaves Mistral's API on 30 Apr 2027 |
Apertus 1.5 (8B, 70B) | ETH Zurich, EPFL and CSCS, Switzerland | 8B and 70B | Apache 2.0 (its makers call it fully open); gated download with an acceptable-use policy | 24 Jul 2026 | About 262k context; image and audio input |
EuroLLM-22B | EuroLLM project | 22.6B | Apache 2.0 | Dec 2025 | Trained on 35 languages, including all 24 official EU languages |
ALIA-40b | Barcelona Supercomputing Center, Spain | 40.4B, dense | Apache 2.0 | 2026 | Pretrained on 35 European languages, tuned mainly for Spanish, Catalan, Basque, Galician and English; about 164k context |
MoE means the model uses only a few expert sub-networks per token: Large 4 works with 49B of its roughly 1T parameters at a time. That makes it fast, but all of it must sit in GPU memory. A dense model uses every parameter for every token.
Mistral is Europe's strongest lab on benchmarks: when it previewed Large 4 in October 2026, Artificial Analysis called it the most intelligent model from outside the US and China. On the current Intelligence Index, v5.0.1, Large 4 (High) scores 26, against a median of 23 for open-weight models of similar size.
Research groups publish open models trained on public supercomputers: Apertus 70B on 4,096 NVIDIA GH200 GPUs, EuroLLM-22B on 400 H100s on MareNostrum 5 in Barcelona. Germany's Teuken-7B is non-commercial only (CC BY-NC 4.0 for version 0.6), so we leave it out.
Which GPU server each model needs
Weights are a model's stored parameters: 2 bytes each at BF16, the usual published format, 1 byte at FP8 and about half a byte at FP4 (NVFP4 is NVIDIA's 4-bit format). Our rule of thumb: keep the weights to about 80% of GPU memory. The rest holds the KV cache, the model's working memory for each conversation; long contexts, many users and fine-tuning need more (see our GPU memory guide).
We arrange whole 8-GPU servers, so the table shows how many copies of each model fit on one. Each copy serves its own users. Memory per GPU: B300 288 GB, B200 180 GB, H200 141 GB, H100 80 GB.
Model, download size | 8x H100 | 8x H200 | 8x B200 | 8x B300 |
|---|---|---|---|---|
Mistral Large 4, FP8, about 1.06 TB | No | No: too little room | 1 copy | 1 copy |
Mistral Large 4, NVFP4, about 0.61 TB | No | Fits, but not advised: NVFP4 is built for Blackwell | 1 copy | 2 copies |
Mistral Large 3, FP8, about 682 GB | 1 copy at NVFP4 (403 GB), per Mistral | 1 copy | 1 copy | 2 copies |
Mistral Medium 3.5, BF16, about 267 GB | 1 copy | 2 copies | 4 copies | 4 copies |
Mistral Small 4, BF16, about 242 GB | 2 copies | 2 copies | 4 copies | 4 copies |
Apertus 1.5 70B, BF16, about 145 GB | 2 copies | 4 copies | 4 copies | 8 copies |
ALIA-40b, BF16, about 81 GB | 4 copies | 8 copies | 8 copies | 8 copies |
EuroLLM-22B or Ministral 3 8B and 14B, BF16, 21 to 45 GB | 8 or more | 8 or more | 8 or more | 8 or more |
FlexiCloud estimates from the published file sizes, with each copy using 1, 2, 4 or 8 of the server's eight GPUs. Mistral's Large 3 card lists FP8 on one server of B200s or H200s and NVFP4 on one of H100s; for Large 4, Mistral says NVFP4 targets Blackwell.
For Large 4, 8x B300 leaves far more room for long conversations (see B300 vs B200). Blackwell GPUs have native FP4 maths; older GPUs can load some NVFP4 checkpoints but run them more slowly (see Hopper vs Blackwell). Splitting one model across eight GPUs works because NVLink joins them inside the server.
Tell us the model, your languages and expected traffic, and we will size the server.
Mistral's licences
Licence | Mistral models | Who it shuts out | What they can do instead |
|---|---|---|---|
Apache 2.0 | Small 4, Ministral 3, Large 3 | Nobody: use it, change it, build it into products or serve it as an API | Keep the licence text and notices with the model |
Mistral Ecosystem Open-Weight Licence (based on MIT) | Large 4 | Any company that, itself or through an affiliate, sells others inference or fine-tuning access with real control over inputs or settings (model-as-a-service). It then needs the licence for any commercial use beyond internal use, not just an API | Ask Mistral for a separate commercial licence |
Modified MIT with a revenue cap | Medium 3.5 | Any company with global consolidated revenue above $20M in the previous month (about $240M a year), and anyone working for one. Fine-tuned and modified versions are covered too | Ask Mistral for a commercial licence (granted at its discretion) or use Mistral's hosted API |
Three common cases for Large 4:
- Running it for your own staff: allowed.
- Building it into your own product's features: allowed. Passing your users' requests to a Large 4 that someone else hosts is not model-as-a-service either.
- Selling API or fine-tuning access, yourself or through an affiliate: you need Mistral's licence before any commercial use of Large 4, or of versions built on it, other than internal use.
Large 4's licence has no revenue test. Medium 3.5 is the opposite: no model-as-a-service clause, but no rights at all above the monthly cap. Both are open weights, not open source: free to download, with conditions.
Medium 3.5 and Large 3 leave Mistral's API on 30 April 2027 (deprecated 31 October 2026), but their weights were still on Hugging Face on 19 November 2026. If a licence pushes you back to an API, see rent, buy or use an API.
This summarises the current licence texts. It is not legal advice.
EU rules and self-hosting
Rule | What it says | If you self-host an open model |
|---|---|---|
AI Act: general-purpose model duties | Apply to model makers since 2 Aug 2025; Commission fines possible from 2 Aug 2026; models released before Aug 2025 have until 2 Aug 2027 | The duties sit with the model maker. You become a modified model's provider only if your changes use over a third of its training compute, which the Commission expects to be rare |
AI Act: high-risk uses | Delayed by the Digital Omnibus on AI, Regulation (EU) 2026/1744: stand-alone uses to 2 Dec 2027, uses built into products to 2 Aug 2028. The general-purpose model rules were not delayed | Depends on what you use the model for, not where it comes from |
GDPR: personal data to the US | The EU-US Data Privacy Framework is in force; the EU General Court upheld it on 3 Sep 2025, and an appeal is pending at the Court of Justice | GPUs in the EU keep prompts in the EU; still check who operates and can access the hardware |
GDPR: personal data to China | No EU adequacy decision (a ruling that a country protects data well enough) for China; Italy's data regulator ordered a block on the DeepSeek app in Jan 2025 | Downloaded weights, even Chinese ones, send nothing to China when they run on GPUs you control |
Who can see the prompts still matters: see bare metal vs cloud GPUs.
Not legal advice.
The UK
The UK is a separate region, outside the EU AI Act.
Topic | Fact |
|---|---|
Home-grown top-tier (frontier) model | None released as of 19 Nov 2026. Google DeepMind is in London but owned by Alphabet, so this series counts its models as American |
Isambard-AI, Bristol | GBP 225M national AI supercomputer: 5,448 NVIDIA GH200 superchips, about 21 AI exaflops; launched Jul 2025 |
UK Compute Roadmap | Up to GBP 2B to 2030, growing the public AI Research Resource 20 times; up to GBP 750M for a national supercomputer in Edinburgh, due 2027 |
AI Growth Zones | Five data-centre zones: Culham (Oxfordshire), North East England, North Wales, South Wales and Lanarkshire (named 29 Jan 2026) |
UK public compute goes mainly to researchers through funding calls, so commercial teams rent their own GPUs. In December 2025 the EU renewed its finding that UK data protection is adequate, until December 2031, so personal data can move from the EU to the UK. A team whose data must stay in the UK should still choose the UK region.
European languages use more tokens
APIs bill per token (see our token pricing guide). A May 2026 preprint averaged six tokenizers, the tools that split text into tokens, on the same texts:
Language | Tokens per word | Compared with English |
|---|---|---|
English | 1.23 | 1.0x |
French | 1.52 | About 1.2x |
German | 1.76 | About 1.4x |
Polish | 2.25 | About 1.8x |
Hungarian | 2.71 | About 2.2x |
Finnish | 2.99 | About 2.4x |
A Finnish or Hungarian word often carries more meaning than an English one, so the gap for the same document is usually smaller. Mistral says its Tekken tokenizer compresses French, German, Italian and Spanish about 30% better than its earlier one (vendor-reported). An August 2026 study of nine models found that, in the median non-English language, models cost 31% more to run than in English and scored about 10% lower. It also found that no open-weight model covers all 24 EU languages well.
On a dedicated server you pay for time, not tokens, so extra tokens mean lower throughput: test tokens per second in your own languages before you size (see GPU cost per hour). Our India and Gulf posts cover Indic languages and Arabic.
EU public compute
Programme | What it is | Where it stands |
|---|---|---|
AI Factories | 19 publicly funded AI supercomputing sites in 16 EU countries, plus 13 smaller access points (Antennas), 6 of them in partner countries including the UK | Selected in 2024 and 2025; none so far in 2026 |
AI Gigafactories | Up to seven sites, described in the 2025 plan as about 100,000 AI chips each; about EUR 10B of public money, plus about EUR 20B hoped for from private investors | Tender opened 30 Jul 2026; bids now due 3 Dec 2026 (large sites) and 4 Feb 2027 (medium), all to be resubmitted. First sites were estimated for mid-2028, before these delays |
France's pledges | EUR 109B of private AI pledges (Feb 2025), then EUR 93B of Choose France pledges (Jun 2026), including one pledge of EUR 45B for AI data centres in northern France by 2031 | Pledges, not built capacity |
These serve research and industrial policy, and the gigafactories are years away, so production teams in 2026 and 2027 rent commercial GPUs.
Five checks before you self-host
- Region: Europe or the UK, set by where your personal data must stay.
- Licence: if you resell model access, Large 4 needs Mistral's licence; above $20M monthly revenue, Medium 3.5 needs one for any use.
- Server size: use the sizing table, with headroom for context and users.
- Language test: tokens per second and answer quality in your languages.
- Contract and access: who sees prompts and logs, data deletion, support hours. See questions to ask a GPU provider.
Need Mistral in Europe or the UK?
Mistral Large 4 at FP8 takes one full B200 or B300 server, and your data rules decide whether it belongs in the EU or the UK. Europe and the UK are two of our six regions. In both, we arrange dedicated 8-GPU B300, B200, H200 and H100 servers from verified operators for your own deployment, on a minimum one-year term billed upfront. We do not host models or sell tokens: you run the model, under its licence.
Get a quote for either region or read why teams rent through us; if you would rather talk, book a call.
Where these figures come from
Checked on 19 November 2026 unless a date is given.
- Mistral: Hugging Face pages for Large 4, Medium 3.5, Large 3 and Ministral 3 8B; Large 4 licence and Medium 3.5 licence; Small 4 launch (16 March 2026); models overview; Tekken tokenizer (July 2024).
- Other models on Hugging Face: Apertus 70B, Apertus 1.5 70B and Apertus 1.5 (CSCS, 24 July 2026), EuroLLM-22B, ALIA-40b, Teuken-7B.
- Benchmarks: Artificial Analysis leaderboard, Large 4 page and preview article (October 2026).
- EU rules: Commission guidelines for general-purpose AI providers; National Law Review on the Digital Omnibus (July 2026); Commission adequacy decisions; WilmerHale on the DPF appeal (December 2025); Euronews on DeepSeek (June 2025).
- UK: UK Compute Roadmap (April 2026), Tom's Hardware on Isambard-AI (July 2025), GOV.UK on AI Growth Zones (29 January 2026), Commission adequacy decisions (UK decisions renewed 19 December 2025).
- Languages: arXiv preprints 2605.24718 (May 2026) and 2608.01395 (August 2026).
- Public compute: EuroHPC on AI Factories (October 2025) and the gigafactory tender; Commission on Antennas and gigafactories; The Next Web (30 July 2026); Reuters on France's 2025 pledges; Euronews on Choose France (1 June 2026).