Mistral Large 4 is open: Europe's largest AI model to download. What it can do and what it needs

TL;DR. On October 29, Mistral released the weights of Mistral Large 4, a model with roughly one trillion parameters, trained in Europe. Any company may download it, run it in-house and adapt it, including for commercial use. At coding it is level with the best open model while using almost half as many tokens. On the Artificial Analysis overall score it reaches 26 points, the top model 59. We measured which machines it fits on. For mid-sized companies, Large 4 is not a new top model but a fallback that no one can switch off anymore. Whether running it yourself pays off depends on your data and your memory, not on the leaderboard.

What Mistral released

Mistral had shown the model as a preview on October 6. After three more weeks of training, the weights have been on Hugging Face since Thursday, with no sign-up required. According to the model card, Large 4 has roughly one trillion parameters, of which only 49 billion are active per token. It reads text and images, handles up to one million tokens at once and covers more than 160 languages, including all official languages of the EU. It was trained on 3,800 NVIDIA chips in Mistral's own data centers in Europe.

In its announcement, Mistral places the model's strengths in coding and cybersecurity. On the Artificial Analysis Coding Agent Index, Large 4 reaches 48%, the best open model GLM-5.3 49%. Large 4 needs 47% fewer tokens per task on average to get there. On Terminal-Bench 4.0, coding in the terminal, its score rose from 28% to 41% since the preview. Through Mistral's API, the model costs $0.70 per million input tokens and $2.10 per million output tokens.

The catch

• The overall score is low. On the Artificial Analysis Intelligence Index, Large 4 scores 26 points. Gemini 4 Carbon stands at 59, Claude Opus 5.5 at 58, the best open models from China at 32. Mistral calls the 26 the highest score among Western open models. That is true, but only just: Meta's Muse Spark Open has 25. Closed models are missing from Mistral's chart.
• Rented, it is no bargain. Artificial Analysis also calculates what one task of the index costs. Large 4 needs $2.06 for 26 points. GPT-6.1 Terra at the "high" setting reaches 42 points for $0.26, DeepSeek V4.1 Flash 32 points for $0.60. If you only rent through an API, you get more for less elsewhere.
• Almost all task scores come from Mistral. A technical report has been announced but is not out yet. So far, mainly the overall score has been measured independently.
• The cyber capability cuts both ways. Mistral itself writes that it sees potential for real attacks and has tightened its safeguards. According to Mistral, the model refuses 96.6% of malicious cyber requests. Only vetted partners get a special version for attack testing. The open weights themselves can no longer be recalled.

We measured: does the model fit in-house?

Open does not mean handy. On October 30, we queried the file sizes of all 18 versions that Mistral and the teams behind llama.cpp and Unsloth have uploaded so far, through the Hugging Face API. Compressed versions store each weight with fewer bits and so save memory.

• The original weighs 1,061 GB in 54 files. Hugging Face serves it on a server with eight H200 cards, 1,128 GB of memory in total. That leaves 67 GB on paper for the conversation history. A single request gets around 65 tokens per second there.
• Mistral's 4-bit version weighs 610 GB and needs NVIDIA's latest Blackwell generation, for example a machine with four B300 cards.
• 8 of the 18 versions stay below 512 GB, all compressed to 3 bits or fewer. 512 GB is what an Apple computer with an Ultra chip has in its largest configuration. Hugging Face names it as the only single machine for the 2-bit versions and measured 15 tokens per second on it. The smallest 4-bit version already needs 605 GB.
• Compression has a cost. Unsloth measured how often the small versions pick the same next token as the original: around 96 times in 100 at 3 bits, around 95 at 2 bits. How that adds up over long tasks is not something this measurement shows.

What the license allows

The license is a modified MIT license. Using, modifying, further training and building the model into your own products is free, including commercially. There is one condition: anyone offering the model to third parties as a service, so that they control inputs, settings or training data themselves, first needs a contract with Mistral. Internal use and products in which the model only sits inside specific features are exempt. A machine builder that makes its service documentation searchable with Large 4 is free to do so. An IT service provider that sells its customers their own Large 4 access is not.

What this means for mid-sized companies

Large 4 does not solve any task that a mid-sized company could not solve more cheaply today through a rented API. Its value lies elsewhere: it is a large model from Europe whose weights no one can take away from you. If certain data must not leave your company, or you do not want to depend on a single provider, you now have a full-size fallback.

For most companies, however, a server with eight data center cards is too big. According to The Next Web, Mistral is planning smaller derived versions, one of them for laptops, but has given no date. Until then, Large 4 mainly matters to data centers, public authorities and large enterprises, and to service providers who run it for others.

What you should do now

1. Settle the obligations first, then the hardware. Write down which data really must not leave your company. Running a model in-house only pays off for those tasks.
2. Build everything else to be swappable. Connect models so that you can switch provider with one setting. Then Large 4 is an option, not a decision for years.
3. Check the license against your case. Internal use and embedded product features are free. If you want to offer the model to customers as their own access, talk to Mistral first.
4. Measure on your own tasks. If you use coding agents, test Large 4 against your current model. Count the cost per completed task, not per million tokens.
5. Watch for the small versions. For a single server, mid-sized open models are more realistic. Once Mistral publishes the announced smaller versions, it is worth taking another look.

Europe now has an open model the size of those from the big labs. Whether it helps you depends on your data and your data center, not on the headline. If you want to work out which of your tasks should run in-house, we can help you choose.