Mirendil unveils Fern, a first model announced with a 10-million-token context
The new lab is betting on context length rather than size. Fern 0.1 is not yet available, and its own coding scores remain well below those of the best models.
Mirendil, a new AI lab, unveiled its first model, Fern, on October 26. Its pitch comes down to one number: a context window announced at up to 10 million tokens, which, according to the company, is about 23 MB of code or half a million lines held in working memory at once. For comparison, OpenAI’s GPT-6.1 Terra, made available the same day on Amazon Bedrock, accepts up to 1 million input tokens.
Betting on context rather than size
Mirendil defends a simple thesis: to make a model more capable, you can increase its parameter count, but also the amount of information it can consult at once. Fern 0.1, the first version, is built on Mirendil-Base, a model of 0.1 to 0.5 trillion parameters (the company only gives a range), pretrained from scratch and then refined with a short round of post-training.
The target uses are those of teams building AI: auditing training data, debugging a training run from its logs, reading agent traces end to end, working on very large code repositories. Two larger base models, Large (0.5 to 2 trillion parameters) and XL (2 to 10 trillion), are in pretraining.
What the company’s numbers show
On context length, Mirendil publishes a RULER test: single-item retrieval (“single needle”) succeeds 100% of the time at every length tested, up to 10 million tokens. But at that length, the score rests on only 20 examples, and the company itself promises more complete results “soon”. Finding a needle in a haystack is the simplest test; reasoning over the whole of such a context is much harder.
In coding, Fern 0.1’s scores, measured by Mirendil with the OpenCode tool in a single attempt, remain far from the best: 73.4% on SWE-Bench Verified and 76.5% on SWE-Bench Pro v2-Hard, where the same chart credits GPT-6 Astra with 90.2% on the latter. Measurement methods differ from one model to another in this chart, which limits the scope of the comparison.
Not usable yet
The key point: Fern 0.1 is not available. The page says access is “coming soon” and offers an early-access request. No price, API or open weights have been announced.
Our take
A 10-million-token context, if truly usable, would change how people work on large codebases or long agent traces. But at this stage, Fern is an announcement: a single context test published on a small sample, coding scores behind the leaders, and no public access. We will come back to it when the model can be tested.
Sources
- Mirendil — Introducing Fern Primary source — scores taken from the page's interactive charts
- AWS — OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock Primary source — point of comparison (GPT-6.1 Terra's context window)
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
