Mistral Large 4: What European AI Teams Should Do Now

Mistral Large 4 is worth evaluating now, through the hosted preview API, but it is not yet a model a European team can plan to run on its own hardware. Mistral promises the weights; as of 27 October 2026 they are not published and no licence has been named. Test it on your own documents and languages today, and treat self-hosting as a decision that waits on the licence text.

That is our reading. The rest of this piece separates what Mistral has reported from what we think it changes for an enterprise team building AI systems in Europe.

What Mistral shipped on 6 October

Mistral announced the model in Introducing Mistral Large 4 on 6 October 2026. According to that post, it is "a 1 trillion-parameter natively multimodal model with 52 billion active parameters", and it "unifies instruction, reasoning, and agentic capabilities in a single model". Mistral says the preview can be tried today through Mistral Studio.

The Mistral Large 4 model documentation gives the exact figures: 1.05T total parameters, 52B active parameters and a 1.6B vision encoder. It lists a 1M-token context window and marks the model as a Public Preview, version v26.10. The same page lists structured outputs, function calling, document Q&A, prefix, chat completions, batching, agents and conversations, and built-in tools among its capabilities. Some early coverage gave a different active-parameter count; we follow the documentation.

The architecture is a mixture of experts (MoE): for each token a router sends work to a subset of the network, so only the active parameters do the arithmetic, while the full set of weights still has to be available to the serving system.

Mistral also makes three claims aimed directly at European buyers. Its announcement says the model "was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe", that "the public preview is served on that same infrastructure", and that "it will be able to run on private cloud or on-premise". It says language coverage spans "more than 160 languages, including every official language of the European Union".

Open weights: promised, not published

The decisive fact for self-hosting is still missing. Mistral's announcement says: "We will release the weights by the end of the month." When we checked on 27 October 2026, the mistralai organisation on Hugging Face showed a Mistral Large 4 collection with no model repository listed, and no licence had been named.

Until the licence text exists, calling Large 4 "open-weight" describes an intention. The licence decides whether a commercial deployment is allowed, under what conditions, and whether a fine-tuned derivative can be shipped inside a product. When it lands, we would read it for three things before planning any hardware:

• Commercial terms. Whether commercial use is permitted outright or conditioned on company size, revenue or a separate agreement.
• Use restrictions. Any acceptable-use list that would rule out a planned workflow, and who decides disputes.
• Redistribution and derivatives. Whether fine-tuned or quantised versions can be deployed to customers or only used internally.

The model supply chain questions apply as usual: where the weights came from, how they are verified, and who is responsible for updating them. This article will need a dated update once the weights appear.

Price and what it means for cost per task

The model documentation lists hosted pricing of $1.36 per 1M input tokens, $0.14 per 1M cached input tokens and $4.18 per 1M output tokens. Because the model is a Preview, prices may change before general availability.

Token prices are only one input to cost per task. What a task actually costs depends on how much context each call carries, how often that context repeats and can be cached, how long the model reasons before answering, how many tool calls an agent makes, and how often a result fails verification and has to be retried. A cheaper model that needs more retries or longer prompts can end up costing more per completed task than a dearer one that gets it right first time. The cached-input rate matters most for workloads that reuse a long fixed context, such as a policy manual or a contract template, across many calls.

The practical step is to measure tokens, retries and pass rates per task on a representative sample, and only then compare models. Our guide to modelling the cost of an AI system covers the rest of the bill: infrastructure, evaluation, monitoring and the people who review outputs.

Reading the benchmarks

Every headline number so far is Mistral's own. The announcement reports that the model "scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4". In Mistral's human coding evaluation, "ML4 Preview ranked second of five models (3.74)", behind Claude Opus 5.

Vendor-reported results are useful for knowing what a vendor optimised for. They are not independent tests, and the settings behind them are the vendor's choice. Second of five in a vendor's own evaluation is an honest-sounding result, but it tells a team little about how the model handles their invoices, their Catalan support tickets or their tool schemas.

The more relevant claim for a European team is the language coverage. If a workload involves documents in several EU languages, that is where a pilot should spend its effort: the same task, run in each language the business actually uses, scored against the same acceptance criteria. Our page on evaluating AI systems describes how we build those test sets.

Sovereignty and self-hosting realities

Mistral's statement that the public preview is served on its European infrastructure is a provider statement, and it is a useful one. It is not the same as a data-processing commitment. Before sending personal or confidential data through the API, a team still needs to check, separately, the region its account and data are actually processed in, the retention period for prompts and outputs, whether data is used for training, and what the contract and data processing agreement say. Our pages on data residency and GDPR and EU data residency list the questions we ask of any provider.

Self-hosting, if the licence permits it, is a capacity question first. In an MoE model the active parameters set the compute per token, but the total weights set how much memory the serving system needs to hold or stream. With 1.05T total parameters, that points to a multi-GPU deployment for most teams. Quantisation reduces the footprint, and offloading some experts to host memory or storage is possible, but each of those trades accuracy or latency for capacity, and the right balance depends on the workload's throughput and response-time needs. We would size it from measured traffic, not from the active-parameter figure alone; GPU sizing and cost explains the method, and on-premise AI covers what running it in-house involves.

There is also a regulatory angle. A model of this scale falls into the general-purpose AI model category under the EU AI Act, and the obligations for the team deploying it depend on what the system built on top of it does, not on the model alone.

What we would do this quarter

Our recommendation for a European enterprise team is short:

1. Pilot through the API on non-sensitive workloads. Pick two or three real tasks, preferably document and multilingual work, and measure pass rate, retries and tokens per completed task against the model currently in use.
2. Put it behind a router. A hybrid model routing layer keeps the choice cheap and reversible: if the Preview changes, or the licence disappoints, traffic moves back without rewriting the application.
3. Wait for the licence before buying hardware. Revisit self-hosting once the weights and licence are published, with the measured traffic from the pilot in hand.

For a wider view of where Mistral fits among European and open-weight options, see our Mistral model page and the comparison of open weights and proprietary models. If a team wants help designing that pilot or the routing layer around it, contact us.