Mirendil has introduced Fern, its first language model, trained from scratch and able to take in up to 10 million tokens at a time. Mirendil, the startup AI Insiders has covered for its $200M seed round and its $100M Google Cloud deal, says that is roughly 23 MB of code, or about half a million lines.
The first release, Fern 0.1, had only light post-training and targets repetitive agentic work: tracing failures in a model’s training logs, auditing datasets, and combing through whole agent transcripts. Mirendil puts it at 0.1 to 0.5 trillion parameters and says it handles text and code. Access is listed as “coming soon.”
Two bigger siblings are in the works. Mirendil says the larger one is more than 80 percent through pretraining and the biggest, at 2 to 10 trillion parameters, is past 20 percent.
Mirendil’s post promises fuller long-context results later and shows only a basic test on a standard benchmark, so the headline window has not been independently checked. Teams planning long-context agents should wait for those numbers before shifting any workloads.
Reported by Mirendil on 26 October 2026.
