Mirendil launched Fern 0.1, its first model, with access currently limited to an early request queue.
The Mirendil announcement positions Fern as a long-context model for auditing training data, debugging training runs from logs, and reading agent traces end to end. Its headline claim is a context window of up to 10 million tokens—about 23 MB of code or roughly 500,000 lines—though Startup Fortune reports that Mirendil’s cited multi-hop accuracy reached 45% at full length.
For builders, Fern could cut some of the chunking and retrieval overhead around large codebases, logs, and traces. But it is still hard to evaluate broadly: access is not public yet, pricing has not been disclosed, and the published performance figures come from Mirendil’s own testing.