Article
million Token Context is only the starting point: Former Anthropic employees want to hand an entire AI laboratory over to Agents
Dongcha Beating AI news flash: Mirendil has released its first model, Fern 0.1, supporting up to 10000000 Token of context, enough to accommodate approximately 500000 lines of code or several days of AI Agent work records at once. But what the company truly wants is for AI to participate in developing the next generation of AI, gradually taking over the work of frontier AI laboratories.
Mirendil made its June public debut this year. Its team of approximately 20 people comes from Anthropic, Google DeepMind, OpenAI, and xAI. Co-founder Harsh Mehta previously built Anthropic's internal automated AI research platform. At its founding, the company raised a a16z seed round led by 200000000 USD and Kleiner Perkins, and subsequently signed a multi-year computing partnership with Google Cloud valued at more than 100000000 USD.
Mirendil aims to bring the automated AI R&D capabilities available internally at major companies to external researchers, enabling Agents to handle code writing, experiment execution, model training, and troubleshooting. Teams researching drugs or new materials can develop models for their specialized problems without first assembling a top-tier AI team.
October 21, Mirendil disclosed an experiment. A researcher, armed with AI Agent, started from an empty code repository and completed pretraining Mirendil-Base from scratch in just 3 and a half months. The training data came from public code, web pages, and datasets including GitHub, with no model distillation or synthetic data used. The Agent wrote more than 99%of code and was also responsible for running experiments and troubleshooting. The researcher primarily set the research direction and screened experimental results. Including early trial and error, data processing, and Agent calls, the computing cost of the entire project was less than 10000000 USD。
5 The Fern 0.1 released days later was built on this base model. It is specifically designed for AI R&D, using ultra-long context to read large codebases, training logs, and Agent execution records. Previously, when this material was too long, it often had to be truncated, summarized, or retrieved first, making it easy to miss critical details. Mirendil wants models to directly read more raw records, investigate why training failed, or review exactly what an Agent did over several consecutive days.
However, 10000000 Token is not the first to emerge. Google tested this length as early as 2024, and Meta's Llama 4 Scout, released last year, also supports ten-million-scale context. The latter was actually pretrained and post-trained using only 256000 Token in length, then extended to 10000000。