Mirendil Launches Fern 0.1 With 10 Million Token Window
The new model offers high-capacity data retrieval to assist researchers in training runs and experiment design.
Updated on Oct. 26, 2026 in Artificial Intelligence
Live Poll
Do you trust that massive AI context windows are currently reliable enough for complex, high-stakes tasks?
Mirendil has released Fern 0.1, an AI model featuring a 10 million token context window. The model, which is currently available via early access request, is designed to assist researchers in designing experiments and evaluating machine learning workflows.
Why it matters
Fern 0.1 demonstrates a shift toward high-capacity context windows developed with relatively low capital investment, potentially accelerating the development of specialized research-grade AI tools.
Fern 0.1 handles a 10 million token context window, achieving a 100% score on single-needle retrieval tests. The foundation model, Mirendil-Base, was constructed in 3.5 months using less than $10 million in training compute.
The players
Mirendil
A startup focused on foundation model development that reached a $1 billion valuation shortly after its launch.
The details
Fern 0.1 is a post-trained version of Mirendil-Base, a foundation model developed by a single researcher. The model utilizes a massive context window—the amount of data an AI processes before generating an output—to manage complex data arrays. While it excels at retrieving specific points of information from large datasets, its performance drops to 45% accuracy when performing multi-hop reasoning, which requires connecting disparate facts to reach a conclusion.
Timeline
June 2026: Mirendil launched with a $200 million seed round.
October 26, 2026: Mirendil released Fern 0.1 and opened access requests.
The Tech Race
Mirendil's release significantly expands the competitive landscape for long-context AI, far exceeding the capacity of existing models like Gemini 1.5 Pro. This launch indicates a move toward highly specialized tools built on lean development cycles.
Researchers and enterprise teams can now request early access to the model to begin testing its long-context capabilities in their workflows. There is currently no public pricing information or broad release date for those outside the initial access group.
The takeaway
Mirendil has proven that massive context windows can be built in months rather than years at a relatively low compute cost. Researchers should monitor the upcoming access rollout to see how the model performs in real-world experimentation versus the reported benchmark figures.
Further reading
For more on the current state of large language models, visit the Artificial Intelligence section.
Live Poll
Do you trust that massive AI context windows are currently reliable enough for complex, high-stakes tasks?
