Mistral Large 4: should you self-host it?
On 6 October 2026, Mistral AI launched a preview of Mistral Large 4, a model of around 1 trillion parameters whose weights it promises for the end of October. European, announced as open weight and ranked by Artificial Analysis as the strongest model from outside the US and China: should you host it yourself? Not before its licence is published, and only on a server with 8 B200 GPUs (in FP8) or 8 H200s (in 4-bit). For most uses, one or two GPUs are enough.
What Mistral announced
Mistral Large 4, released as a preview on 6 October 2026, is Mistral AI's largest model to date (the company itself nicknames it "le Chonk"). It is a mixture of experts (MoE): the model is split into specialised sub-networks, and only a few of them work on each token (a fragment of a word, the basic unit of text for a model). It can answer directly or reason at greater length before answering. It reads text and images, and replies in text. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres in Europe. The preview is available through the API in Mistral Studio.
Before publishing the weights, Mistral is having the model tested by cybersecurity leaders, vetted partners and state authorities, who get access to a version with reduced moderation and expanded cyber capabilities. The public API, for its part, remains filtered.
- Size: around 1.05 trillion parameters according to the documentation. Only about fifty billion of them work on each token produced (49 billion, or 52 billion including the input and output layers, according to Mistral's Hugging Face page). On top of that comes a vision module of 1.6 billion parameters.
- Context (the amount of text the model can take into account at once): 1 million tokens according to the documentation, around half a million as measured by Artificial Analysis on the preview API.
- Languages: more than 160 according to Mistral, including every official language of the European Union.
- API pricing: a full price of $1.36 per million input tokens and $4.18 per million output tokens, 2.7 to 2.8 times the price of Large 3. Mistral is halving that price for the two weeks following the launch on 6 October 2026 ($0.68 and $2.09 shown in its documentation on 17 October).
A word of caution on scores. Mistral's own figures are provisional: the last training phase (reinforcement learning, which refines the answers) is still under way. As far as we know, only one independent evaluation has been published, by Artificial Analysis. On its composite index (the Artificial Analysis Intelligence Index), the preview scored 38 points in October 2026, the best score outside the US and China. According to The Decoder, the leader of the same ranking at the same date, Claude Opus 5.5, scored 58.
Announced, not yet verifiable
Date of the weights: Mistral promises them "by the end of the month". AFP, VentureBeat and Journal du Net cite 27 October, Blog du Modérateur 26 October, and Mistral's Hugging Face page 31 October.
Licence: none has been published. VentureBeat mentions a custom Mistral licence, without confirmation from the company.
Hardware: Journal du Net and Blog du Modérateur report compressed versions (FP8 and FP4) running on 4 to 8 NVIDIA B200 or B300 GPUs, the most recent generation. Neither the official announcement nor the documentation mentions them.
What it changes for sovereignty
For a public body or a European company, the appeal fits in one sentence: a model of this calibre, designed and trained in Europe, whose weights are due to become downloadable. Mistral says it will then be able to run in a private cloud or on premises. Your documents would no longer have to leave your perimeter, which no API, even a European one, can offer.
Until then, the preview is served from Mistral's European data centres, according to the company. It has also announced, without a date, a European deployment of its API that it will operate end to end, independently of other providers. That option is a serious one, but your data still leaves your information system. Depending on how sensitive the data is and on your obligations (for instance a SecNumCloud requirement, the French security qualification for cloud services), the difference can be decisive (see our page on AI sovereignty).
Then there is the licence. "Open weight" means downloadable, not free for any use. Even within Mistral, licences vary: Large 3 and Small 4 are under Apache 2.0, a permissive licence that allows commercial use; Medium 3.5 is under a modified MIT licence that excludes companies whose consolidated monthly revenue exceeds $20 million. For Large 4, no licence has been published. Until it is, we advise against any production project.
How many GPUs does it take to run?
You need around 1,240 GB of GPU memory in FP8, which means a server with 8 B200s, and around 700 GB in 4-bit, which means a server with 8 H200s (PixelRaise estimate as of 17 October 2026, ahead of the official technical specifications). The reason lies in the architecture: in an MoE, each token activates only a few experts, but the next token may call on others. To serve the model at a reasonable speed, every parameter therefore has to stay in the memory of the graphics cards (the VRAM). Memory depends on the total (1.05 trillion parameters), compute on the active parameters (around 50 billion): per token, Large 4 costs as much to compute as a dense model of around 50 billion parameters, but it has to be housed like a very large one.
Precision is the number of bits used to store each parameter: 16 in BF16 (2 bytes), 8 in FP8 (1 byte), and 4 in a 4-bit format (half a byte, plus small scaling factors). Hence 2,100, 1,050 and around 590 GB of weights. Reducing precision is called quantisation: less memory, with a risk of lower quality. Here is what the method behind our VRAM calculator produces.
Mistral Large 4: memory and GPUs required (estimate)
| Precision | Weights only | Total memory | H100 80 GB | H200 141 GB | B200 180 GB | B300 288 GB |
|---|---|---|---|---|---|---|
| BF16 (16-bit) | 2,100 GB | ~2,480 GB | 35 GPUs (5 servers) | 20 GPUs (3 servers) | 16 GPUs (2 servers) | 10 GPUs (2 servers) |
| FP8 (8-bit) | 1,050 GB | ~1,240 GB | 18 GPUs (3 servers) | 10 GPUs (2 servers) | 8 GPUs (1 server) | 5 GPUs (1 server) |
| 4-bit (NVFP4) | ~590 GB | ~700 GB | 10 GPUs (2 servers) | 6 GPUs (1 server) | 5 GPUs (1 server) | 3 GPUs (1 server) |
| For comparison: Mistral Small 3.2 24B in BF16 | 48 GB | ~57 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
In FP8, Large 3 fitted on a single server with 8 H200s, according to its Hugging Face page. Large 4, more than half as large again (1.05 trillion parameters against 675 billion), should no longer fit there by our estimate: its weights alone (1,050 GB) would fit in the server's 1,128 GB, but the runtime overhead would not. In 4-bit, that server is enough, although the effect of quantisation on quality has yet to be measured. Finally, every conversation in progress takes up working memory, the context cache (KV cache): on 8 B200s in FP8, only around 46 GB is left for it. By our estimate, that represents between 130,000 and 650,000 tokens to share among all simultaneous users, depending on the architecture Mistral publishes. The "4 to 8 B200 or B300 GPUs" reported in the press are consistent with these orders of magnitude.
An NVIDIA DGX B200 server takes up 10U (rack units) and draws up to around 14.3 kW. To rent, at public prices on 17 October 2026, budget €42 excluding VAT per hour at OVHcloud for 8 H200s (enough in 4-bit) and €60 excluding VAT per hour at Scaleway for 8 B300s (enough in FP8). Rented around the clock (730 hours a month), such a server therefore costs roughly €31,000 to €44,000 excluding VAT per month. In its cloudgouv region, which holds the SecNumCloud qualification from ANSSI, the French national cybersecurity agency, Outscale lists €6.24 excluding VAT per hour per H200 GPU, or around €50 excluding VAT per hour for 8 GPUs, not counting processors and memory. That is in a different league from the cost of an internal AI running on one or two GPUs.
To experiment without GPUs of this calibre, open source tools such as llama.cpp or KTransformers can keep the experts in the server's main memory (RAM). KTransformers reports around ten tokens per second, a few words per second, for Kimi K2 (1 trillion parameters, 32 billion active) with around 600 GB of RAM and a GPU with around 14 GB of memory. Large 4 uses more parameters for each token (49 billion), so it would probably be slower. Enough to test, not to serve an organisation.
When to choose it, and in what form
Large 4 makes sense where your own tests show a clear gain. Mistral highlights it for code, cybersecurity and automation, based on its own evaluations. There are four ways to access it, from the simplest to the most demanding.
- Mistral's API. The preview is available today, paid per token, with no infrastructure. Mistral has announced a European deployment, without a date. Your data leaves your information system: keep this route for the data your risk assessment allows out.
- A European cloud, once the weights are published. A rented 8-GPU server on which you run the model yourself. If you need a SecNumCloud-qualified region: at Outscale, whose most powerful GPU in that region is the H200, plan for one server with 8 H200s in 4-bit, or two servers in FP8. The data stays in an environment you control, with a European host, but the meter runs full time.
- A shared computing centre. For local authorities, hospitals or universities, it is a realistic way to make such a server pay for itself: several organisations share the infrastructure and the model under joint governance. A sovereign option, but the slowest to organise.
- Your own servers. The option that keeps everything in-house, provided your server room can take 10U, 14 kW machines and your teams know how to run them.
When to stay with a model that fits on 1 or 2 GPUs
For most internal uses (document search, summarising, drafting, answering the public), a much smaller model is enough. Mistral Small 3.2 (24 billion parameters) fits on one 80 GB GPU. Mistral Small 4 (119 billion, under Apache 2.0) needs around 144 GB in FP8 by the same method: two 80 GB GPUs are only just enough, with no margin for simultaneous users, while two H200s leave room. Mistral itself lists a minimum infrastructure of 4 HGX H100, 2 HGX H200 or one DGX B200. And perceived quality depends above all on the document retrieval chain: that is the subject of our article on sovereign RAG.
Start from the model your hardware can serve to all your simultaneous users with acceptable latency. On an open weight stack, the model remains a replaceable component: if a specific need justifies Large 4 later, it can slot into the same chain (same documents, same retrieval) without migrating your data, once inference engines support it. You will, however, need the matching hardware.
What to check when the weights are published
Publication of the weights, expected at the end of October 2026, will settle four questions.
- The licence. Apache 2.0, modified MIT or a custom licence? Check commercial use, use by a public authority, fine-tuning (retraining on your data) and redistribution, for the exact version you will deploy.
- The technical specifications. Architecture (which determines the working memory per user), official compressed versions, recommended hardware, compatibility with common inference software (vLLM, llama.cpp). It will confirm or correct our table.
- The published version. Does it include the end of the reinforcement learning run? Is it the public API model, with its filters, or the reduced-moderation version?
- Independent evaluations. Further measurements on the final model and, above all, your own test on your own documents.
Key takeaways
Mistral Large 4 is good news for sovereignty: the European model ranked highest by Artificial Analysis should become downloadable by the end of October 2026. But self-hosting it requires a server with 8 high-end GPUs (H200s in 4-bit, B200s or B300s in FP8), and therefore a data centre, a budget and a team. Our recommendation comes in three steps:
- Now: keep a model on one or two GPUs for everyday uses, and note the tasks where it reaches its limits.
- When the weights are published: read the licence and the technical specifications. No production deployment until the licence is compatible with your use.
- Then: test Large 4 on those tasks only, with your documents, through the API if your data allows it, otherwise on a server rented in Europe or a shared one.
Frequently asked questions
Should you self-host Mistral Large 4?
Not yet. As of 17 October 2026, neither the weights nor the licence have been published. After that, you will need a server with 8 high-end GPUs (8 B200s in FP8, or 8 H200s in 4-bit), and therefore a data centre, a budget and a team. For most uses, a model on one or two GPUs is enough; test Large 4 first on the tasks where that model reaches its limits.
Is Mistral Large 4 open source?
Mistral presents it as an open weight model and promises its weights for the end of October 2026. As of 17 October 2026, neither the weights nor the licence have been published, and Artificial Analysis still lists it as proprietary. Open weight does not mean open source: what you may do with it will depend on the licence.
How many GPUs do you need to self-host Mistral Large 4?
By our estimate of 17 October 2026 (using the method of our VRAM calculator), you need around 1,240 GB of GPU memory in FP8, which means one server with 8 B200 GPUs or two servers with 8 H200s, and around 700 GB in 4-bit (NVFP4 format), which means one server with 8 H200s. Mistral has not yet published an official configuration.
When will the Mistral Large 4 weights be available?
In its announcement of 6 October 2026, Mistral only promises them "by the end of the month", meaning the end of October 2026. AFP, VentureBeat and Journal du Net cite 27 October, Blog du Modérateur 26 October, and Mistral's Hugging Face page shows an expected date of 31 October.
Should you replace Mistral Small with Mistral Large 4?
Not by default. For document search, summarising or drafting, a model that fits on one or two GPUs remains the most economical choice. Large 4 is justified for complex tasks where your own tests show a gain, if you have the necessary infrastructure.
Which model for your uses and your hardware?
A pilot lets you test, on your own documents, a model sized for your infrastructure, before deciding whether to aim bigger.
Reading
Sources
- Mistral AI: "Introducing Mistral Large 4" (official announcement, 6 October 2026)
- Mistral AI: Mistral Large 4 model card (documentation, parameters, context, pricing)
- Mistral AI: changelog (preview and launch discount)
- Mistral AI on Hugging Face: Mistral-Large-4-1T-A52B (preview page, expected weights date)
- Artificial Analysis: evaluation of the Mistral Large 4 preview
- Artificial Analysis: Mistral Large 4 Preview model page (weights status, measured context)
- The Decoder: Mistral Large 4 against the leading models
- VentureBeat: Mistral Large 4 launch and open weights timetable
- Alliancy (with AFP): Mistral Large 4, open weight and cyber defence (in French)
- TechCrunch: Mistral's new 1 trillion parameter model
- Journal du Net: Mistral Large 4, FP8 and FP4 versions and weights date (in French)
- Blog du Modérateur: Mistral Large 4 and Chinese open weight models (in French)
- IT Social: Large 4 moves upmarket and up in price (in French)
- Mistral AI: Mistral Large 3 model card (documentation, pricing)
- Mistral AI: Mistral 3 announcement (Large 3, Apache 2.0 licence)
- Mistral AI on Hugging Face: Mistral Large 3 675B Instruct (FP8 deployment)
- Mistral AI: Mistral Small 4 announcement (Apache 2.0 licence, minimum infrastructure)
- Mistral AI on Hugging Face: modified MIT licence of Mistral Medium 3.5
- Mistral AI on Hugging Face: Mistral Small 3.2 24B Instruct
- Hugging Face: "Mixture of Experts Explained"
- NVIDIA: DGX B200 datasheet (memory, rack space, power draw)
- NVIDIA: H200 datasheet
- NVIDIA: H100 datasheet
- NVIDIA: introducing the NVFP4 format
- OVHcloud: Public Cloud pricing (GPU instances)
- Scaleway: GPU instance pricing
- Outscale: public pricing (cloudgouv-eu-west-1 region)
- Outscale: support FAQ on the SecNumCloud qualification of the cloudgouv-eu-west-1 region
- KTransformers: guide to running Kimi K2
- Moonshot AI on Hugging Face: Kimi K2 Instruct (total and active parameters)
- llama.cpp: server documentation (--cpu-moe and --n-cpu-moe options)