Mirendil announces Fern, its first model, with up to 10M tokens of context

Mirendil, a new AI lab, announced Fern on October 26, 2026: a model it says it pretrained from scratch, with support for up to 10 million tokens of context. Mirendil says that is about 0.5 million lines of code (roughly 23 MB), and that prediction keeps improving up to 10M tokens.

Mirendil's claims (not independently verified):

• Fern 0.1 is Mirendil-Base after a short round of post-training. Mirendil shared agentic benchmark results with the announcement.
• It is built for high-volume agentic work, such as auditing training data, debugging training runs from logs and reading agent traces end to end.
• Fern is the first and smallest of three models; two larger ones are being pretrained now.

The model is not yet generally available: Mirendil says it will make it available soon and takes early-access requests on its site. The announcement came days after Mirendil's post "Pretraining at scale without a pretraining team", which describes how it can pretrain large models.