Mirendil released Fern, its first model. It is pretrained from scratch. Fern supports a context window of up to 10 million tokens. Its first version, Fern 0.1, completed a short post-training round. New versions will ship frequently. Fern targets high-volume agentic tasks. It can hold 23 MB of code or half a million lines at once. It is built for auditing training data, debugging runs, and analyzing full agent traces. Early access is available by request. A larger model, Mirendil-Base-Large, is 80%+ through pretraining. An even larger model, Mirendil-Base-XL, is 20%+ through pretraining.
What changed
No prior Mirendil model existed.
What it unlocks
Running agentic tasks across entire large codebases, logs, and traces without truncation.
- 10M token context window
- 0.1–0.5T parameters (Fern 0.1)
- 23 MB of code capacity
- 500,000 lines of code capacity
What you need to act on it
- early access request
Sources