Mirendil unveils fern 0.1, a code-focused AI model with 10m-token context window

Mirendil, an AI R&D lab, has launched Fern 0.1, its first AI model, designed for AI research and development rather than general conversation. Fern is a coding model with a context window of up to 10 million tokens, equivalent to about 23 MB of code or roughly half a million lines. This places it on par with Megatron-LM in terms of scale. The model was built from Mirendil-Base after a short post-training phase. Training data came exclusively from public sources, including GitHub, web pages, and public datasets, without using distillation, synthetic data, or licensed data. The code was developed from scratch using PyTorch and JAX. Fern is the smallest of three planned models, with Mirendil-Base-Large and Mirend, which are in advanced stages of pretraining. Mirendil claims that AI agents wrote over 99% of the code, with a single researcher overseeing the process. The model outperformed several competitors in benchmark tests, including Inkling and Nemotron 3 Ultra, though it lagged behind some frontier models in specific tasks.