Large 2 Is Mistral’s New Language Model. It’s a European Ode to Efficiency

Mistral AI, the promising French artificial intelligence startup backed by Microsoft, has announced Large 2, its new and most powerful language model to date. It aims to compete with the latest releases from Meta and OpenAI.

Why it matters. This announcement demonstrates the rapid evolution of Mistral, a French national jewel only a year old that represents Europe's hope in AI. Also, it challenges the much more established and capitalized U.S. companies in the field.

The details:

• Architecture: Large 2 has 123 billion parameters, less than a third of Meta’s Llama 3.1 (405 billion), launched a day earlier.
• Significant context: It offers a 128,000 token window, equivalent to about 300 pages of text, matching competitors such as GPT-4.
• Performance: According to synthetic tests, it outperforms models such as Claude 3.5 Sonnet and approaches GPT-4o performance in coding or mathematical tasks.
• Versatility: It can generate code in more than 80 different languages with high accuracy.
• Polyglot: it supports dozens of languages, including French, German, Italian, Spanish, Portuguese, Arabic, Russian, Chinese, Japanese, and Korean.

Focus on efficiency: Mistral has put a lot of emphasis on efficiency of Large 2. Despite its reduced size compared to Llama 3.1, it achieves similar or superior performance in several metrics.

This emphasis can translate into lower implementation and operational costs for end users.

Performance compared to other models. Large 2, like Llama 3.1 (not available in Europe), is only partially open source in the traditional sense. You can try it without a license for non-commercial and research purposes. Still, any commercial use requires a license fee. This hybrid approach is becoming the norm for AI models.

Availability and access. Large 2 is available through the company’s ‘la Plateforme’ under the “mistral-large-2407” name. You can also access it via corporate clouds: Google Cloud (Vertex AI), AWS (Bedrock), Azure (AI Studio), and IBM (watsonx.ai).

Those interested can also experiment with the free trial of the model on ‘le Chat,’ the ChatGPT competitor developed by Mistral.

Implications. Mistral’s rapid progress intensifies competition, improves efficiency, and helps democratize advanced AI available on multiple platforms.

It also puts Europe on the AI map, which the U.S. and China dominate.

This article was written by Javier Lacort and originally published in Spanish on Xataka.

Image | Mistral, Mockuuups Studio

Related | The EU’s Regulatory Obsession Is Creating a World Where AI Will Advance at Two Speeds—and Europe Will End Up Losing