Mistral Large

Mistral AI has launched Mistral, its latest and most advanced large language model (LLM). This model excels in multiple languages, reasoning, math, and generating code. You can access Mistral Large through their platform called la Plataforme and on Microsoft Azure. It’s also available for testing in their new chat app, le Chat.

A comparison chart shows how Mistral Large stacks up against other powerful LLMs, like GPT-4 and Gemini Pro. It ranks second after GPT-4 on the MMLU benchmark, achieving a score of 81.2%.

Mistral Large Performance

Mistral Large Capabilities  
Mistral Large has several impressive features:

- A context window of 32,000 tokens
- Fluent in multiple languages, including English, French, Spanish, German, and Italian
- Strong abilities in reasoning, knowledge, math, and coding tasks
- Supports function calling and JSON format natively
- A smaller version called Mistral Small was also released for low-latency applications
- Allows developers to create moderation policies with precise instructions

Reasoning and Knowledge 
The following table shows how Mistral Large performs on reasoning and knowledge benchmarks. While it is behind GPT-4, it still outperforms other models like Claude 2 and Gemini Pro 1.0.

Mistral Large Performance

Math and Code Generation 
The table below displays Mistral Large's performance in math and coding benchmarks. It shows strong results in math and GSM8K tasks, but it falls short in coding benchmarks compared to models like Gemini Pro and GPT-4.

Mistral Large Performance

Multilingual Abilities
The table below illustrates Mistral Large's performance in multilingual reasoning benchmarks. It outperforms Mixtral 8x7B and Llama 2 70B in all tested languages, including French, German, Spanish, and Italian.

Mistral Large Performance

Mistral Small
Alongside Mistral Large, Mistral AI has also introduced a smaller and optimized model called **Mistral Small**. This model is designed for low-latency tasks and surpasses Mixtral 8x7B in performance. Mistral AI reports that Mistral Small has strong capabilities in RAG-enablement, function calling, and JSON format.