TL;DR. With the Mistral 3 family of December 2025, Mistral switched to Apache 2.0. Mistral Large 3 and the small Ministral 3 models may run in-house commercially and free of charge. On October 29, 2026, Mistral Large 4 followed with open weights, though under a modified MIT licence. So the licence is hardly the hurdle any more. The bill is: through Mistral’s own API, 50 million tokens a month cost about $49; a rented server with eight H200 cards about $31,500. Running it yourself therefore pays off almost only when regulation or data protection rule out the cloud. Top performance remains with the closed models.
Three questions: when does open source pay off?
Choosing between an open model and a frontier API is not a matter of belief. It comes down to three variables: regulatory pressure, token volume and the complexity of your agents. Overlook one and you build the wrong stack. The questions are deliberately blunt, because grey areas lead to hybrid setups that are expensive to untangle later.
| Question | If yes |
|---|---|
| 1 · Regulation: medical devices class IIa and up, defence, critical infrastructure, core banking systems? | Self-hosting is mandatory, for example with Mistral Large 3 or 4 or a Ministral 3 model |
| 2 · Volume: permanently very high load in the range of many billions of tokens a month? | Calculate self-hosting; only here can it pay for itself |
| 3 · Agents: long tool chains, screen control, real-time speech? | Frontier model; open models lag further behind here |
| Otherwise: knowledge base, classification, extraction, drafting, coding help | API with EU hosting, or self-hosted right away if your data demands it |
The myth that “we have to host everything ourselves” rarely survives the calculation. What has changed with Mistral 3 and Large 4: those who do pass the test now have strong models they are allowed to use for it.
Mistral 3 and Large 4 versus frontier: the situation in November 2026
The first version of this post in May still compared Mistral Large 2 under a restrictive research licence. That world has been gone since December 2, 2025. The Mistral 3 family is entirely under Apache 2.0, and since late October the new flagship has been open as well:
| Model | Size | Licence |
|---|---|---|
| Mistral Large 4 | about 1 trillion parameters, 49 billion active per token, text and images, 1 million tokens of context | modified MIT licence |
| Mistral Large 3 | 675 billion parameters, 41 billion active | Apache 2.0 |
| Mistral Medium 3.5 | 128 billion parameters | modified MIT licence |
| Ministral 3 | 3, 8 and 14 billion parameters, each also as a reasoning variant, with image understanding | Apache 2.0 |
The modified MIT licence of Large 4 allows use, modification and further training, commercially too. There is one condition: if you offer the model to third parties as a service of your own, you need a contract with Mistral first. Internal use stays free. A machine builder who makes its service documentation searchable with it is free to do so. An IT provider selling customers its own Large 4 access is not.
Two honest caveats. First: “one trillion parameters” is a headline number. What matters in operation is that only 49 billion compute per token. The memory, however, has to hold all the weights. Second: Mistral does not play at the top. In the Artificial Analysis Intelligence Index, Large 4 scored 26 points in late October, the best closed models 58 and 59.
Artificial Analysis Intelligence Index, as of October 30, 2026
closedopen weights
The decisive question for mid-sized companies is not “best model in the world?” but: is it good enough for knowledge bases, extraction and coding help? Since Mistral 3, the answer there is almost always yes. Then “good enough, own infrastructure, clear GDPR position” beats the last point on the leaderboard.
And the rest of the open-weight competition? For Llama 4, the use policy excludes companies based in the EU from the multimodal models, a real stumbling block. Qwen (Apache 2.0) and DeepSeek (MIT) are free in licence terms and technically strong. With public authorities and critical infrastructure, though, the question of the model’s origin comes up regularly. OpenAI’s gpt-oss (Apache 2.0) is a legitimate option from the US that you can run yourself. Mistral’s distinction is the combination of an open licence and a vendor that is itself based in the EU and subject to the GDPR.
Total cost of ownership: what self-hosting costs
The cost calculation decides the architecture, not the vendors’ slides. Since Mistral 3 it is no longer a licensing question; open weights carry no fee per token. What remains is hardware, power and operations. We calculate with two documented prices:
- API: Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens at Mistral. With 80% input and 20% output, that averages $0.98 per million tokens.
- Own server: Large 4 runs at full precision on a server with eight H200 cards. Rented, one H200 costs $5.40 per hour at Nebius, so the whole server in continuous operation comes to about $31,500 a month.
Monthly costs for Mistral Large 4: API versus own server
50M tokens a month
5B tokens a month
32B tokens a month
Mistral’s APIrented server, continuous operation
At a realistic mid-sized load of 50 million tokens a month, the API is therefore about 640 times cheaper. Break-even only comes at around 32 billion tokens a month, and that does not yet say whether a single server can even handle that volume. Purchased hardware calculates differently, but needs staff, power and cooling that are missing from this calculation. On top of that, a mid-sized company’s load rarely runs around the clock. A server that sits idle at night and at weekends still costs money.
Hybrid setups are therefore more honest than pure self-hosting: sensitive tasks in-house, everything else and every peak load through an API with EU hosting. How far API prices have fallen is shown in our price comparison of frontier providers.
Hardware: what runs on what
If you have to self-host despite this calculation, for example because of regulation or IP protection, choose by model class:
- Single card and edge (Ministral 3 with 3, 8 or 14 billion parameters). From small industrial PCs to workstations to a single 24 GB card, the largest model in a compressed version. Classification, extraction and small knowledge bases are possible without a server room project. This is the underrated entry class.
- One server (Large 4). The original weighs 1,061 GB and runs on eight H200 cards with 1,128 GB of memory in total. Mistral’s 4-bit version weighs 610 GB and needs NVIDIA’s latest Blackwell generation, for example a machine with four B300 cards. For many simultaneous users, inference servers such as vLLM or SGLang are part of the setup.
- Not advisable: cluster plans without a dedicated operations team. A cleanly set-up single server beats a half-finished pair of servers in practice. And the Ministral class covers more everyday tasks than most expect.
EU sovereignty in 2026: from positioning to infrastructure
In 2025 the EU argument was mainly legal. Mistral is a French company and directly subject to the GDPR. With self-hosting or hosting in the EU, the constructs for data transfers to the US fall away. Not because US vendors are bad, but because the legal situation becomes simpler when data and vendor sit in the same jurisdiction.
In 2026 this has turned into infrastructure. Mistral has borrowed $830 million to build a data centre south of Paris with 13,800 NVIDIA GB300 chips. By the end of 2027, 200 megawatts of computing capacity are to be built in Europe. According to its model card, Large 4 was already trained on 3,800 NVIDIA chips in Mistral’s own data centres in Europe. The semiconductor equipment maker ASML led a €1.7 billion funding round in September 2025 and invested €1.3 billion itself. For the self-hosting decision this means the EU option is no longer a niche gamble; it has industrial backing.
What does not run in-house at the end of 2026
Three capabilities remain ahead with the closed vendors for now. If you need them, you cannot work fully in-house:
- Real-time speech. The end-to-end audio systems of the big vendors answer on the phone without a noticeable pause. Open chains assembled from speech recognition, language model and speech output are mostly still too slow for that.
- Screen and browser agents. Models trained to operate user interfaces have a training data lead at the big vendors.
- Top-level multimodality. Image understanding has arrived in open models; Large 4 and Ministral 3 read images. Real-time speech-to-speech and deeply integrated multimodal agents have not.
The question is not whether open models will catch up, but when. If you build an architecture for three years today, plan this gap explicitly instead of overlooking it.
FAQ: Mistral on-prem in practice
May I run Mistral models commercially in-house?
Yes. Large 3 and the Ministral 3 models are under Apache 2.0: commercial use, modification and self-hosting without a fee per token. Large 4 and Medium 3.5 are under a modified MIT licence. Internal use is free; if you offer Large 4 to third parties as a service, you need a contract.
What hardware does Mistral Large 4 need?
At full precision, a server with eight H200 cards; in Mistral’s 4-bit version, about four B300 cards. If you want to start smaller, take a Ministral 3 model that runs on a single card.
Why not simply Llama 4?
Because of the licence. The use policy excludes companies based in the EU from the multimodal Llama 4 models. If you want legal clarity and image understanding, Apache 2.0 models are the simpler route.
Is Mistral as good as GPT, Claude or Gemini?
At the top, no. In the Artificial Analysis index, Large 4 stood at 26 points in late October, the best closed models at 58 and 59. For internal knowledge bases, extraction and coding help, the gap usually does not matter in practice. There, control over the data decides, not the last point.
When should I stay with the API?
When neither regulation nor volume forces you to self-host. Then use an API with EU hosting and postpone the decision. The open weights are not going anywhere. That is precisely the value of the open licence: the path to running it yourself stays open at any time, without dependence on the vendor.
As of November 3, 2026. If you want to check whether your use case needs self-hosting or whether an API with EU hosting is enough, we are happy to talk.
