Mistral Large

Informazioni su Mistral AI

Fondato 2023 · Paris, France

Mistral AI was founded in April 2023 in Paris by Arthur Mensch (former Google DeepMind research scientist, now CEO), Guillaume Lample (former Meta FAIR researcher, co-author of LLaMA) and Timothee Lacroix. The company is widely regarded as Europe's flagship frontier AI lab and has championed an open-weights strategy alongside premier commercial models. Mistral became one of the fastest companies to reach unicorn status, raising EUR 105M in seed funding in June 2023 followed by EUR 385M Series A in December 2023 and a EUR 600M Series B in June 2024 at a EUR 5.8B valuation. Investors include Andreessen Horowitz, Lightspeed, General Catalyst, Bpifrance, BNP Paribas, NVIDIA and Salesforce. Mistral has released Mistral 7B (Sep 2023), Mixtral 8x7B and 8x22B Mixture-of-Experts models (Dec 2023 / Apr 2024), the Mistral Large flagship (Feb 2024) and Mistral Large 2 (124B params, July 2024). The company also ships Codestral for code, Pixtral for vision, Mistral Nemo (in partnership with NVIDIA) and Le Chat, a consumer assistant. Partnerships include Microsoft Azure, AWS Bedrock, Vertex AI and a strategic deal with the French government.

Architettura

Decoder-only Transformer (dense, multilingual flagship)

Mistral Large is Mistral AI's flagship dense Transformer family. The original Mistral Large launched in February 2024 with 32K context. Mistral Large 2 followed in July 2024 with 123B parameters, a 128K context window and significantly improved code and reasoning capability. It is a decoder-only Transformer pretrained on a multi-trillion-token corpus emphasising high-quality multilingual web text, code repositories (80+ programming languages), books and scientific papers. Mistral has been a vocal advocate for European data and language coverage, with first-class performance on French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese and Korean. Post-training applies supervised fine-tuning followed by Direct Preference Optimisation (DPO) on a high-quality, multi-domain preference dataset. The model supports function calling, JSON mode, parallel tool calls and structured output. Mistral Large 2 weights are published under the Mistral Research License for non-commercial research, with a separate commercial license required for production deployment. Mistral also makes the model available through Le Plateforme (their API), Azure AI Studio, Amazon Bedrock, Google Vertex AI, IBM watsonx and Snowflake Cortex.

Parametri

123B (Mistral Large 2; original Mistral Large undisclosed)

Contesto

128.000 token

Capacità

• 128K token context window
• Strong multilingual performance: French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean
• Code generation across 80+ programming languages
• Function calling and parallel tool calls
• JSON mode and Structured Output
• Strong math reasoning on MATH and GSM8K
• Available via Le Plateforme, Azure, Bedrock, Vertex AI, IBM watsonx and Snowflake
• Weights downloadable under Mistral Research License
• Open ecosystem with vLLM, llama.cpp, HuggingFace support
• European data residency option for regulated industries
• Best for: enterprise European deployments, multilingual chat, coding assistants, GDPR-sensitive workloads.

Addestramento e licenza

Multi-trillion-token pretraining mixture of curated multilingual web text (heavy emphasis on European languages), code repositories spanning 80+ languages, scientific papers and books. Knowledge cutoff is approximately mid-2024 for Mistral Large 2.

Licenza: Dual license: Mistral Research License (non-commercial research) for downloadable weights; Mistral Commercial License required for production use. Also available via API and cloud-marketplace endpoints.

Test di sicurezza: Mistral applies supervised safety training and DPO with safety preferences. The company's stated philosophy favours steerability and minimal default refusals; third-party evaluations have noted lower refusal rates than US frontier labs.