Mistral Large

What is Mistral Large?

Mistral Large is a highly optimized large language model developed by Mistral AI. It is part of a suite of models that includes Mistral Small, which is optimized for latency and cost. Mistral Large incorporates innovations such as RAG-enablement and function calling, similar to Mistral Small. These models are part of Mistral AI's simplified endpoint offering, which aims to provide a range of solutions from lightweight to flagship models.

Mistral Large is a text generation model built for complex multilingual reasoning tasks such as text understanding, transformation, and code generation. At its February 2024 launch, Mistral AI positioned it as the second-best model available via API, behind GPT-4, based on contemporary benchmark results; the broader LLM landscape has since moved on considerably.

Key features of Mistral Large include:

• Multilingual fluency in English, French, Spanish, German, and Italian, offering a deep understanding of grammar and cultural nuances.
• A 32K token context window for accurate information retrieval from extensive documents.
• Advanced instruction-following capabilities that allow for custom moderation policy design, demonstrated in the system-level moderation of le Chat.
• Native function calling ability, combined with a constrained output mode on la Plateforme, facilitating large-scale application development and technological stack modernization.

Mistral Large Capabilities

At launch, Mistral AI evaluated Mistral Large against the era's leading Large Language Models (LLMs) across a variety of benchmarks covering reasoning, knowledge, multilingual tasks, math, and coding.

In reasoning and knowledge, Mistral AI reported that Mistral Large outperformed other pre-trained models of the time in benchmarks such as MMLU, HellaSwag, Wino Grande, Arc Challenge, TriviaQA, and TruthfulQA.

On multilingual capabilities, Mistral AI reported that Mistral Large exceeded the performance of LLaMA 2 70B in benchmarks conducted in French, German, Spanish, and Italian, at that time (Figure 3).

In math and coding, Mistral AI reported strong launch-era performance across popular benchmarks including HumanEval, MBPP, and GSM8K, relative to models available in February 2024.

These results, reported by Mistral AI at launch, showcased the model's broad and versatile capabilities across a wide range of tasks and languages relative to models available at the time.

Function calling and JSON format

Function calling allows developers to connect Mistral endpoints to their own tools for advanced interactions with internal code, APIs, or databases.

The JSON format mode ensures language model outputs are in valid JSON, facilitating easier integration and manipulation of data within developers' pipelines.

At launch, these features were exclusive to Mistral Small and Mistral Large.

Partnership with Microsoft

At Mistral Large's launch, Microsoft and Mistral AI announced a multi-year partnership to bring Mistral's models to Azure, alongside a minority investment by Microsoft in Mistral AI. Mistral Large was the first Mistral model made available through Azure AI Studio.

How to Access Mistral Large?

There are three ways to access Mistral offerings:

• La Plateforme — Hosted on Mistral's secure European infrastructure, this platform allows developers to build applications using Mistral's model library.

• Azure Integration — Access Mistral Large via Azure AI Studio and Azure Machine Learning. At launch, Mistral AI reported early success with Azure beta testers.

• Self-deployment — For highly sensitive applications, deploy Mistral models in your own environment. This option includes access to the model weights.

Mistral Large Use Cases

Mistral Large model excels in a wide range of NLP tasks, including:

• Code Debugging — Mistral Large supports developers with precise code suggestions and error identification, using its reasoning capabilities to analyze code logic and pinpoint issues.
• Translation & Summarization — Provides robust support for translating multiple languages and efficiently condenses text to highlight key points.
• Text Generation & Classification — Capable of generating coherent text from prompts and accurately categorizing content into specific groups.
• Customer Service Excellence — Mistral Large drives superior customer service by offering immediate, precise, and tailored responses via automated chatbots, boosting customer satisfaction and engagement through quick and effective query resolution.
• Content Creation & Education — Enhances marketing content creation and automates educational processes, including grading and learning enhancements.

These applications underscore Mistral Large's capability to handle diverse tasks involving text understanding, processing, generation, and user interaction across multiple domains.