Mistral is releasing Magistral Small as a 24-billion-parameter open-source model under the Apache 2.0 license, while Magistral Medium remains proprietary and available only in preview mode12. The open-source version addresses what Anjney Midha from Andreessen Horowitz, who led Mistral's Series A funding round, calls the key advantage of open-source alternatives to "closed, centralized AI systems"3.

Magistral Medium claims performance of 1,000 tokens per second, which Mistral says outpaces rival systems3. The models are designed for enterprise applications including "structured calculations and programmatic logic to decision trees and rule-based systems," according to the company's blog post2.

The launch comes as reasoning models gain traction across the industry, with companies investing in systems that use additional computing power during inference to improve accuracy1. Mistral's Magistral directly competes with OpenAI's o1 model and China's DeepSeek R1, both released in recent months2.

Unlike competitors that primarily reason in English or Chinese, Mistral emphasizes multilingual capabilities, particularly for European languages. "Historically, we've seen US models reason in English and Chinese models reason in Chinese," Mensch said. "We're trying to push reasoning capabilities in European languages"2.

Founded in April 2023, Mistral has achieved rapid growth, reaching over $100 million in contracted sales just 15 months after launching its first paid products, according to a source familiar with its finances1. The company, valued at $6.2 billion as of June 2024, generated $30 million in revenue in 2024 and projects $60 million for 20252.

The models are being deployed across Mistral's Le Chat platform and select APIs, with broader access expected following the announcement3.