Mistral Large 2
by Mistral AI·Europe·Released
Mistral's flagship — 123B parameters, EU-resident, with research-only weights available.
About this model
Mistral Large 2 (July 2024) is the French lab's flagship — 123B parameters, EU-resident hosting, and competitive with GPT-4o / Claude 3.5 Sonnet on most benchmarks at launch. The weights are available under the Mistral Research License (non-commercial use; commercial requires a paid API or contract).
Mistral's positioning has always been 'the European frontier lab.' For EU enterprise customers dealing with GDPR and the upcoming AI Act compliance, Mistral is often the only frontier-class model that survives procurement.
Strengths
- •Best frontier model for EU AI Act and GDPR compliance
- •Strong multilingual: French, German, Italian, Spanish at parity with English
- •Weights available under Research License — easier than fully-closed competitors
- •EU-resident hosting via Mistral's la Plateforme
Limitations
- •Trails Claude 4 / GPT-4.1 on hardest English-language benchmarks
- •128K context — smaller than top US frontier models
- •Limited multimodal (text + code only; Pixtral Large is separate)
When to use it
- →European enterprise deployments under GDPR / AI Act
- →Multilingual European customer support
- →Public-sector AI procurement where data sovereignty matters
- →Hybrid open + closed model architectures
Architecture & training
123B-parameter dense transformer (Mistral has moved away from MoE for their flagship after Mixtral experiments). Trained on Mistral's GPU infrastructure in France. The accompanying Pixtral Large extends the family with vision support, sharing the same backbone architecture.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MATH | 71.5 | |
| MMLU | 84.0 | |
| HumanEval | 92.0 |