Mistral AI released Mistral Large 2, a multilingual, tool use-capable, open model of its own

Mistral AI has been on a model-releasing roll since last week when it unveiled Mathstral, Codestral Mamba, and Mistral NeMo. Now, just a day after Meta announced Llama 3.1 450B, Mistral AI released Mistral Large 2, a 123B parameter model, under its proprietary research license. Mistral Large 2 enjoys some features that have become standard, such as a 128K token context window, multilingual support (which includes French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean), coding proficiency in 80+ languages (including Python, Java, C, C++, JavaScript, and Bash), and function calling capabilities.

Mistral Large 2 performs remarkably in general knowledge, scoring 84% in the MMLU benchmark. As a point of comparison, Mistral Large 2 closes in on GPT-4, surpasses Claude 3.5 Sonnet, and is less than 4 percent points behind GPT-4o and Llama 3.1 450B. The model also holds its own in coding accuracy, accomplishing results comparable to GPT-4o's and consistently outperforming Llama 3.1 450B in the MultiPL-E benchmark tests. Mistral AI has improved Mistral Large 2's instruction-following and alignment capabilities (compared to its predecessor). In addition to being less prone to hallucination, the model is better at recognizing that it cannot provide a complete reliable answer to a query. Finally, Mistral Large 2 can handle parallel and sequential function calls, making it ideal for complex enterprise use cases.

Mistral Large 2 is available on Mistral's Le Plateforme and powering the Le Chat assistant. The weights for the instruction-tuned model can be downloaded from Hugging Face. The model is also available with Mistral's cloud partners: Google Vertex AI, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai.