Mirendil has introduced Fern, its first model, pretrained from scratch and built for high-volume agentic tasks. Its headline specification is a context window of up to 10 million tokens, though access is listed as coming soon.
Mirendil Watch analysis
What happened
The company describes Fern 0.1 as its first version, completed with a short round of post-training. Its launch page says the model accepts text and code and lists a size range of 0.1–0.5 trillion parameters. Mirendil says new versions will ship frequently.
The company’s Fern announcement does not give a release date for access or explain how performance varies across the listed size range. Those details matter: a huge context window is a specification, not a demonstration of how well a model handles long tasks.
Why it matters
A 10-million-token context window would let a model take in unusually large amounts of text or code at once. For agentic work, that could mean more room for extensive project material or long-running tasks, rather than repeatedly trimming context to fit. The practical value will depend on access, speed, reliability and cost, none of which the announcement specifies.
Fern also gives Mirendil a named first model and a stated direction: focused, high-volume agentic tasks. That makes this more than a vague promise to enter the AI market, while leaving the most useful questions for hands-on testing.
Our read
The scale is striking; the evidence so far is a launch-page specification, not a performance result. If you work with long documents or codebases, Fern is worth keeping an eye on, but there is little to evaluate until access opens and the company publishes concrete usage details. A very long context window is an invitation to test what fits, not proof that the model can use it all well.
What to watch
- When early access opens, and whether it is limited to selected users.
- How the different model sizes compare on long-context and agentic tasks.
- Pricing, speed and any limits on the advertised context window.
Discussion spark: Would a 10-million-token context window change how you use an AI model, or would you wait for evidence that it handles long tasks reliably?
Sources and evidence
- Source update (26 October 2026, 00:00 UTC)
not affiliated with or endorsed by Mirendil