Mirendil is selling a different thesis: stop trying to cram knowledge into weights and just give the model room to read. Its first release, Fern 0.1, accepts contexts as long as 10M tokens, which the lab translates into about 23 MB of source text, roughly half a million lines, or a codebase on the order of Megatron-LM resident in memory at once.
The founding story doubles as marketing. According to Mirendil, one researcher pretrained the base model over 3.5 months for under $10M of total compute, a figure that folds in experimentation, agent token spend and data processing, with AI agents producing more than 99 percent of the code. Training corpora were public only: no distillation, no synthetic examples, no licensing deals. The stack was written from scratch on PyTorch and JAX.
Long-context retrieval is where it shines. In base form Fern hits 100 percent on RULER single-needle recall at every window from 32K through 10M tokens. Reasoning is another matter. Terminal Bench 4.0 comes in at 6.1 percent, against 66.4 percent for Claude Opus 5.5 in Mirendil’s own comparison table, and multi-hop queries fall to 45 percent at full length.
Against shipped frontier systems the raw quality picture is unflattering. At a 1M-byte window Kimi-K3 posted 0.140 bits per byte while Fern managed 0.165, so Fern trails on equal-context prediction. What it offers instead is a curve that keeps gaining where competitors have already plateaued.
Fern is the first of three planned models, and the smallest. Mirendil-Base-Large is past 80 percent of pretraining, with Mirendil-Base-XL beyond 20 percent. There are no open weights; access is by early-release request.
