A one-year-old artificial intelligence lab has made a claim that, if it holds up in real use, changes what a single AI model can hold in its head. Mirendil released Fern 0.1 on October 26, a model with a context window of 10 million tokens. The company says that is roughly 23 megabytes of code, or about 500,000 lines. For comparison, Google's newer Gemini 4 Argon and Anthropic's Claude Opus 5 both offer 1-million-token windows. Fern's figure is ten times larger.

A context window is the amount of text a model can consider at once. The most useful work in software and research rarely fits in a short prompt: an entire codebase, a full contract, or days of an AI agent's logs all exceed today's mainstream limits. The usual workaround is retrieval — chop the material into pieces and feed only the most relevant chunks to the model. That can miss the one file nobody thought to retrieve. A 10-million-token window promises to make that workaround unnecessary for certain jobs.

Built for researchers, not chatbots

Mirendil is not chasing the consumer chatbot market. It launched out of stealth in June with a $200 million seed round at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia also investing. Its founders previously worked at Anthropic and Google, and its mission is to build AI that helps researchers design experiments, tune training runs, and evaluate models faster than a human team can.

Fern fits that mission directly. Mirendil says Fern 0.1 is a post-trained version of Mirendil-Base, a foundation model it says was built in 3.5 months by a single researcher, for under $10 million in compute. AI agents wrote more than 99% of the code, ran experiments, and diagnosed failures, while the researcher decided what to try, what to trust, and when to intervene. All training data came from public sources, the company says, with no distillation, no synthetic data, and no purchased data.

The benchmark behind the headline

The number worth reading past the headline comes from Mirendil's own evaluation. It tested the base model with RULER, a long-context benchmark developed by Nvidia researchers, at the full 10-million-token length. Based on 20 examples for each reported result, it reported 100% accuracy on single-needle retrieval, 90% on multi-key retrieval, and 45% on multi-hop questions.

That split is the honest core of the story. Finding one fact buried in a huge document is far easier than chaining several facts together to answer a hard question. A 45% score on multi-hop questions at full length means the model gets those wrong more than half the time. For a lab pitching itself on rigour, publishing that number rather than burying it counts in its favour.

What happens next

Access to Fern is listed as coming soon, with Mirendil taking early access requests and no pricing disclosed yet. Enterprise clients and research laboratories are expected first, not the general public. The company is pitching Fern at auditing training data, debugging failed training runs from logs, and reading an AI agent's full multi-day trace end to end — tasks where a huge, literal window beats retrieval because nothing can afford to be missed.

The open question is whether a 10-million-token window becomes a durable advantage or a temporary lead. Gemini and Claude could simply extend their own windows past 1 million tokens, and a bigger window is only useful if the reasoning inside it keeps up. If Fern proves reliable in real research workflows, it will show that small, agent-assisted teams can build tools the giants have not.