Mistral AI is a French company formed as Mistral AI SAS, founded in 2023 in Paris by Arthur Mensch (formerly of Google DeepMind), Guillaume Lample, and Timothée Lacroix (both from Meta AI and graduates of École Polytechnique). The name references the mistral, a powerful northwesterly wind that blows along France's Mediterranean coast. From its inception, the company has focused on developing large language models, including open-weight LLMs, and quickly became one of the key European players in this field.
From the start, Mistral AI has demonstrated a strategy of releasing models with diverse architectures and purposes, effectively implementing cutting-edge scientific advancements. Its first release, the Mistral 7B model in 2023, uses Sliding Window Attention and Grouped-Query Attention, which significantly accelerated inference and ensured GPU memory savings. Continuing its development of architectural innovations, the company introduced Mixtral 8x7B—the first production-grade model based on a Sparse Mixture of Experts (SMoE). Another direction involved working with larger models and multimodality: Mistral Large 2 uses 123 billion parameters, and its training substantially increased the proportion of non-English data, enabling it to handle over 12 natural languages and 85+ programming languages. Pixtral Large became a native multimodal model with a unique image processing mechanism. In parallel, Mistral experiments with fundamentally new paradigms, such as implementing a State Space Model architecture in Codestral. The Mistral 3 series is highly varied—released under the open Apache 2.0 license are both giant multimodal MoE models like Mistral Large 3 with native FP8 and NVFP4 quantization, and lighter models with 3, 8, and 14 billion parameters, available in reasoning and instruct versions.
In addition to open models, the company maintains the API platform La Plateforme, the chat interface Le Chat, and official SDKs, thereby covering the full spectrum of LLM use cases—from research experiments to premium deployment and corporate SaaS solutions.
Thus, Mistral AI has become one of the symbols of the "European school" of open LLMs, consistently publishing model weights under Apache 2.0 or its own research licenses and making them available for deployment in nearly all modern frameworks. This combination of open models and a commercial API solidifies Mistral's role as one of the key hubs for the open-source LLM community.