French AI startup Mistral released its new flagship model Large 2 on July 24, shortly after Meta's release of Llama 3.1.
Mistral AI, founded in June 2023, is one of the best-funded and most competitive AI companies in Europe. Its core members come from top AI institutions like Google DeepMind, and co-founder Mensch is the author of papers on large models such as Chinchilla, proposing core technologies including Scaling Laws.
The company secured 105 million euros in funding just 4 weeks after its establishment, focusing on open-source large model development. In December 2023, Mistral released the Mistral 8x7B open-source large model with 56 billion parameters, comparable in efficiency and performance to LLaMA-65B. The company also launched a multilingual conversational assistant called Le Chat.
The Large 2 model uses 123 billion parameters, about one-third of Llama 3.1's 405B parameters, but surpasses the latter in code generation, mathematics, and reasoning abilities. Large 2 generates more concise responses, avoiding lengthy descriptions, offering cost advantages and faster local running speeds.
Although Large 2 lacks multimodal capabilities, it can rival Llama in accuracy and reliability of conversational responses. The model's training process focused on addressing hallucination issues and improved instruction following and multi-turn conversations.
Large 2 has a 128k context length and supports multiple languages and 80 coding languages. The model is available on Google, Amazon, Azure, and IBM platforms. Users can experience it on Mistral's La Plateforme or test it for free on Le Chat.
Mistral completed its Series B funding in June this year, raising $640 million at a $6 billion valuation. The company currently has about 60 employees, most of whom are engaged in product development and research.