Mistral AI

Based in France, Mistral AI develops fine-tuned language models and provides businesses with foundational technologies to automate existing processes and develop new products. 

Mistral AI has developed a diverse range of AI models to address various language and multimodal tasks. The company expanded its capabilities with Mistral Large series for advanced multilingual reasoning. Mixtral 8x22B, a 176 billion parameter mixture-of-experts model, was released for general language tasks. For specialized applications, Mistral AI introduced Codestral for multi-language code generation, Codestral Mamba for faster code generation using the Mamba architecture, and Mathstral for mathematical reasoning. Mistral NeMo, a 12 billion parameter model, caters to multilingual tasks with a 128k token context length. In the multimodal domain, the company launched Pixtral 12B and Pixtral Large for image understanding and text generation tasks. These models are available through various channels, including free downloads and commercial solutions on major cloud platforms, with an additional content moderation API for their Le Chat service. Mistral Small (22B parameters, 128k tokens) is optimized for high-volume, low-latency tasks like translation and summarization, offering cost-effective performance. Ministral 3B (3B parameters, 128k tokens) is designed for smaller devices like smartphones, excelling in efficiency for lightweight applications. Ministral 8B (8B parameters, 128k tokens) targets more powerful devices like laptops, balancing performance and resource requirements for on-the-go AI workloads.

Mistral 7B uses GQA, which combines multi-query and multi-head attention. This results in faster inference and the ability to handle longer sequences efficiently.

Key customers and partnerships

In December 2023, Mistral AI partnered with Google Cloud to distribute its LLMs globally using Google's AI-optimized infrastructure for developers and businesses.In January 2024, the company partnered with Brave, a provider of a browser-based AI chatbot named Leo, to introduce the updated Mistral AI's Mixtral 8x7B as the default LLM. Mistral AI formed several partnerships in February 2024 to introduce its foundation models to its partners' platform. These included Capgemini, a provider of technology and consulting, Amazon, Microsoft, and IBM to integrate Mistral 7B and Mixtral 8x7B into Amazon Bedrock, Microsoft's Azure, and watsonx cloud platforms, respectively. In March 2024, the company partnered with Snowflake, a data management and analytics platform, to provide Mistral Large in the Snowflake data cloud. In April 2024, Mistral partnered with Amazon to introduce Mistral Large to Amazon Bedrock. In June 2024, the company partnered with SAP to offer access to its latest models.

In July 2024, the company partnered with Dassault Systemes to expand the latter’s product portfolio and with BNP Paribas, a global bank to provide its large language models (LLMs).