Mistral Large 2 and Mistral Small 3.1: the EU model pair with FR/DE/IT strength

What is Mistral Large 2?

Mistral Large 2 is the flagship language model from French Mistral AI, originally released in July 2024 and available as of May 2026 in version 24.11 (with updates in February and May 2026). It is a dense transformer with 123 billion parameters and a 128,000-token context window. Licence: Mistral Research License for self-host (non-commercial), Mistral Commercial License for commercial use – both accessible via mistral.ai.

Mistral AI was founded in 2023 in Paris and as of May 2026 is a key player in the European AI market. Unlike US models (the current top GPT model, the current top Claude model) and Chinese models (DeepSeek, Qwen), Mistral is EU-headquartered in Paris, with data centre partnerships in France and Germany and an explicit EU AI Act compliance strategy. For Swiss setups with EU data residency requirements, a direct argument.

The Mistral model family is broad as of May 2026. Alongside Mistral Large 2 (123B, premium) and Mistral Small 3.1 (24B, mid-tier), there are Codestral 25 (24B, code-specialised, NVIDIA partnership), Pixtral Large (multimodal, image + text), Ministral 8B (edge variant) and several reasoning-specialised variants. As of May 2026 a new Mistral Large generation is announced but not yet released – updates to existing models, however, are regular.

Mistral Small 3.1 (March 2026, Apache 2.0 licence!) is an important variant. With 24 billion parameters and 128k context, it is the most-deployed open-weight variant for EU setups. Performance: near GPT-4o-Mini at markedly lower cost and full self-host option. Mistral Small 3.1 is Apache 2.0 – the cleanest licence for commercial self-host setups in the Mistral family.

Availability May 2026: La Plateforme (Mistral's own API with France hosting), Azure AI Foundry (France Central, Sweden Central regions), AWS Bedrock (Frankfurt, Paris regions), Google Vertex (europe-west region). Self-host via Hugging Face (mistralai/Mistral-Large-Instruct-2411, mistralai/Mistral-Small-3.1-Instruct).

Why Mistral matters for Swiss data

Mistral is the natural first choice in May 2026 for Swiss setups with EU data residency requirements and FR/DE/IT language needs – for five concrete reasons.

First: EU origin and EU AI Act compliance. As an EU company, Mistral has integrated the strict EU data protection and AI Act duties into its product strategy. Model cards, technical documentation and training data descriptions follow EU standards. For the DPIA per GDPR Art. 35 / FADP Art. 22, the template is simpler than for US models where a TIA for third-country transfer adds up.

Second: language strength on EU languages. Mistral trained its model with a significant share of French, German, Italian and Spanish. On MGSM (multilingual math) and on FRA legal benchmarks, Mistral Large 2 is the productively best option among open-weight models as of May 2026. For Swiss fiduciary offices with Romandie exposure or Ticino mandates, Mistral is typically the right choice – Apertus is on par for FR/IT, but on FR-specific legal text every nuance counts.

Third: two-model strategy is operationally simple. Mistral Small 3.1 for standard workloads (classification, triage, simple generation), Mistral Large 2 for complex reasoning cases. Both through the same API. Routing through LiteLLM. Operational coherence is higher than a mix of different vendors.

Fourth: hosting options in the EU are plentiful. La Plateforme in Paris region (France), Azure in France Central or Sweden Central, AWS Bedrock in Frankfurt or Paris, self-host in a Swiss data centre at Infomaniak or own rack. This variety allows a precisely matching architecture – from "all data in CH rack" to "EU cloud with SLA".

Fifth: Mistral Small 3.1 is Apache 2.0. That is the cleanest licence configuration: self-host without commercial restrictions, no research clause, no 700M MAU cap. Attractive for SME setups with clear licence discipline.

Limits: Romansh and Schwizerdütsch are not trained. Apertus remains the right choice here. Complex reasoning at top frontier level (the current top GPT model, the current top Claude model) still slightly exceeds Mistral Large 2 – the gap has narrowed in May 2026.

Mistral in practice

Architecture. Mistral Large 2 is a dense transformer-decoder model with grouped-query attention and sliding-window attention for efficient long-context handling. 123B parameters, 128k context. Mistral Small 3.1 is also dense (no MoE), 24B parameters, 128k context. Both support tool use in the OpenAI schema, structured JSON output, and function calling.

Setup example via La Plateforme. Direct API access:

```python from mistralai import Mistral import os client = Mistral(api_key=os.environ["MISTRAL_API_KEY"]) response = client.chat.complete( model="mistral-large-latest", messages=[{"role": "user", "content": "Summarise FADP Art. 6."}], ) print(response.choices[0].message.content) ```

La Plateforme prices May 2026: around EUR 2-3 per 1M input tokens and EUR 6-9 per 1M output tokens for Mistral Large 2. Mistral Small 3.1 is markedly cheaper: EUR 0.1-0.2 per 1M input.

Self-host example with vLLM. Mistral Small 3.1 on an RTX 4090 24GB:

``` docker run --gpus all -p 8000:8000 \ vllm/vllm-openai:v0.6.3 \ --model mistralai/Mistral-Small-3.1-Instruct \ --max-model-len 32768 \ --quantization awq \ --gpu-memory-utilization 0.92 ```

For Mistral Large 2 (123B), two H100 80GB with tensor-parallel and AWQ 4-bit quantisation are needed.

Azure AI Foundry. In France Central or Sweden Central region via an Azure OpenAI-compatible endpoint wrapper. Advantage: Microsoft EU contract, standard contractual clauses per EU model template, SOC2 / ISO 27001 certifications in place. Cost per 1M tokens comparable to La Plateforme.

AWS Bedrock. In Frankfurt (eu-central-1) or Paris (eu-west-3) region via the bedrock-runtime API with IAM authentication. AWS DPA and EU model standard contract available.

Routing architecture May 2026. A typical Swiss multi-provider strategy looks like this: LiteLLM proxy as front-end. Three providers registered: (a) "mistral-large-eu" via La Plateforme, (b) "mistral-small-local" via vLLM self-host, (c) "apertus-70b" via Swisscom API or self-host. Routing rules by request type: simple classification to Mistral Small (cost-optimal), complex FR/DE/IT reasoning to Mistral Large 2, RM/Schwizerdütsch or maximum sovereignty to Apertus.

Vision capability via Pixtral Large. Mistral released Pixtral Large, a vision-language model – for Swiss receipt photo workflows and contract scanning, an option next to Llama 4 Scout.

Mistral to production in 5 steps

1. 01Use-case analysis: which language mix (FR/DE/IT/EN), which workload class (classification, reasoning, long context), which data residency requirement (EU or strict CH)?
2. 02Model choice: Mistral Small 3.1 for standard workloads and cost-sensitivity, Mistral Large 2 for complex reasoning or multilingual precision.
3. 03Hosting path: La Plateforme (fast, FR hosting), Azure France/Sweden (enterprise contracts), AWS Bedrock Frankfurt (existing AWS infrastructure), self-host via vLLM (full sovereignty).
4. 04LiteLLM wiring: register Mistral as a provider, logical model names (mistral-large-eu, mistral-small-local), routing rules by request type and sensitivity.
5. 05Use-case benchmark: compare 50-100 real CH-typical queries across Mistral, Apertus and Claude, measure hit rate, derive final routing configuration.

When to use Mistral

Mistral is the right choice when (a) FR/DE/IT language strength is needed (Romandie mandates, Ticino cases, EU-wide fiduciary work), (b) EU data residency with clear EU contracts is required, or (c) a cost-optimal open-weight variant with Apache 2.0 licence is sought (Mistral Small 3.1).

Concrete cases: fiduciary office in Geneva with FR-speaking client communication – Mistral Large 2 via La Plateforme. Law firm in Lugano with IT-FR-DE mixed mandate – Mistral Large 2 for complex cases, Mistral Small 3.1 for standard triage, both through LiteLLM routing. EU-wide consulting boutique with strict GDPR client data policy – Mistral via Azure France Central with EU model standard contract.

For cost-optimal self-host setups in Swiss SMEs, Mistral Small 3.1 is particularly attractive. Apache 2.0 licence, 24B parameters, runs on a single RTX 4090 or L40S, performance close to GPT-4o-Mini, fully EU-compliant. A fiduciary with ten staff can run Mistral Small 3.1 on its own workstation and cover 80 percent of standard workloads with it.

When not to use

For Romansh and Schwizerdütsch, Apertus is the natural choice, not Mistral. Mistral's training data has captured CH-specific language variants only marginally.

For setups with maximum CH data sovereignty (all data must stay in Switzerland, no abroad hosting), Mistral is only an option as self-host – La Plateforme and Azure France hosting are EU but not CH. Apertus with Swisscom API or self-host remains the first choice here.

For top frontier reasoning (math olympiad level, hard legal argument chain with three to four steps), the current top Claude model or the current top GPT model are still slightly ahead. Mistral Large 2 is close but not at the peak.

For long-context workloads beyond 128k tokens, Llama 4 Scout with 10M is the right choice, not Mistral.

For pure hobby exploration without productive ambition, the Mistral Research License clause for Mistral Large 2 self-host is a caveat – Mistral Small 3.1 (Apache 2.0) resolves that for SME workloads.

Trade-offs

FAQ

Related topics

Sources

1. Mistral AI – official models and pricing · 2026-05
2. Mistral Large 2 – Hugging Face model card · 2026-05
3. Mistral Small 3.1 – Apache 2.0 release notes · 2026-03
4. Mistral La Plateforme – API documentation · 2026-05