According to CoinMeta and OneMillion.AI, the Fern 0.1 model released by Mirendil supports up to 10 million token contexts, and is capable of accommodating approximately 500,000 lines of code or several days' worth of AI agent work records. The company hopes to involve AI in the development of the next generation of AI, gradually taking over the work of the cutting-edge laboratory. The Mirendil team, consisting of about 20 people from Anthropic, Google DeepMind, OpenAI, and xAI, has a co-founder Harsh Mehta who previously built the automated AI research platform for Anthropic. The company received a $200 million seed round led by A16Z and Kleiner Perkins, and signed a computing power collaboration worth over $100 million with Google Cloud. On October 21st, Mirendil demonstrated an experiment where researchers collaborated with AI and agent to start from a blank code repository and complete the pre-training of Mirendil-base in just 3 and a half months, with computational costs of less than 10 million US dollars. The Fern 0.1 model focuses on AI development, utilizing ultra-long contexts to read large codebases and training logs.