The latest release from AI startup Mirendil is a coding model capable of processing roughly 500,000 lines of code in a single context window. Fern 0.1, the first model from the company, supports up to 10 million tokens, allowing it to ingest approximately 23 megabytes of data at once for high-volume tasks like auditing training data and reading agent traces end to end.
The system is a post-trained version of Mirendil-Base, a foundation model the company built in 3.5 months with one researcher and less than $10M in compute. Mirendil is currently pretraining two larger models to follow Fern, while its Mirendil-Base-Large model previously recorded the lowest perplexity of all models tested on private codebases by Baseten.