A million-token context is just the starting point: former Anthropic employees want to hand the entire AI lab over to Agents.

Beating AI News Flash: Mirendil has released its first model, Fern 0.1, supporting a context of up to 10 million tokens, capable of holding approximately 500,000 lines of code at once, or several days of AI Agent work records. But what the company really wants to do is have AI participate in developing the next generation of AI, gradually taking over the work of frontier AI labs.

Mirendil made its public debut in June this year, with a team of about 20 people from Anthropic, Google DeepMind, OpenAI, and xAI. Co-founder Harsh Mehta previously built Anthropic's internal automated AI research platform. The company secured a $200 million seed round led by a16z and Kleiner Perkins at inception, and subsequently signed a multi-year compute partnership with Google Cloud worth over $100 million.

Mirendil hopes to bring the automated AI R&D capabilities found inside big tech companies to external researchers, letting Agents handle code writing, experiment running, model training, and troubleshooting. Teams researching drugs or new materials no longer need to first assemble a top-tier AI team to have the opportunity to develop models for their domain-specific problems.

On October 21, Mirendil published an experiment. A single researcher working with an AI Agent, starting from an empty code repository, completed the from-scratch pretraining of Mirendil-Base in just 3.5 months. Training data came from public code such as GitHub, web pages, and datasets, without using model distillation or synthetic data. The Agent wrote over 99% of the code and was also responsible for running experiments and troubleshooting. The researcher mainly decided research directions and screened experiment results. Including early trial and error, data processing, and Agent calls, the entire project's compute cost was under $10 million.

Fern 0.1, released 5 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. In the past, when such content was too long, it often needed to be truncated, summarized, or retrieved first, making it easy to miss key details. Mirendil wants the model to directly read more raw records, trace why training failed, or review what an Agent actually did over several consecutive days.

However, 10 million tokens is not the first occurrence. Google tested this length as early as 2024, and Meta's Llama 4 Scout released last year also supports ten-million-level context. The latter actually used only 256,000 token length for pretraining and post-training, then extended to 10 million through length generalization.

A longer window also does not mean the model can understand all the information within it. Mirendil's tests show that as context grows to 10 million tokens, the base model's code prediction error continues to decline. But in the RULER test, the accuracy of finding a single piece of information from 10 million tokens reached 100%, while when required to connect multiple clues to answer a question, it was only 45%. Moreover, only 20 questions were tested per item at that length. How Fern actually handles such long inputs, how much GPU memory is required, and the speed and cost of full input have not yet been disclosed in detail by the company.

Mirendil is already training two larger models, with the largest planned to reach 2 trillion to 10 trillion parameters. The company also plans to have its own models participate in the development of next-generation models, taking over more work originally done by human researchers. Currently, Fern 0.1 is not yet open for use and is only available through early access applications.