Mirendil releases Fern 0.1, citing context up to 10 million tokens

Mirendil calls Fern its first model and says it was pretrained from scratch.[2] The company describes Fern 0.1 as Mirendil-Base after a short round of post-training.[2] It lists the model at 0.1 trillion to 0.5 trillion parameters and lists text and code as its input types.[2] Mirendil says Fern supports a context window of up to 10 million tokens and describes it as its most efficient model, built for focused, high-volume agentic tasks.[2] Those specifications make the model’s stated context capacity and intended workload central points for teams to assess.

Startup Fortune reported Mirendil results of 100% accuracy on single-needle retrieval and 90% on multi-key retrieval.[1] It also reported 45% accuracy on multi-hop questions.[1] These results cover different retrieval tasks, so the figures should not be treated as a single overall performance score. Startup Fortune said each RULER result reported at 10 million tokens was based on 20 examples.[1] It also reported that Mirendil evaluated the base model with RULER at the full 10-million-token length.[1] The reported sample size is a useful caveat for teams considering how much weight to give the long-context results.

Mirendil says its first pretrained model was built by one researcher in 3.5 months, using less than $10 million in compute.[3] The company says AI agents wrote more than 99% of the code and that it did not purchase or license additional data.[3] Mirendil lists auditing training data, debugging training runs from logs and reading agent traces end to end as use cases.[2] Those examples point to tasks involving long text or code inputs, but the company’s stated use cases are not evidence that Fern has been tested successfully in every such workflow. Teams can use them to frame their own evaluations.

Mirendil lists Fern’s access status as coming soon.[2] Startup Fortune reported that Gemini 4 Argon and Claude Opus 5 have one-million-token context windows.[1] That comparison puts Fern’s stated maximum context length at a different scale, but it does not establish comparative performance or suitability for a particular task. Mirendil says its mission is to accelerate science and technology by democratizing frontier AI research and development.[4] It also said Andreessen Horowitz and Kleiner Perkins led its $200 million seed round, followed by investment from NVIDIA and others.[4] For prospective users, availability and task-specific testing remain practical next steps.