Early preview release of Fern 0.1 by @mirendil. Built on Mirendil-Base, the 0.1–0.5T coding model one researcher pretrained from scratch in 3.5 months for under $10M, with agents writing over 99% of the code. Target use: high-volume work for people building AI, like auditing training data, debugging runs from logs and reading agent traces end to end. Main feature: 10M-token context, about 23 MB of code or ~500K lines, roughly a Megatron-LM-sized codebase. - Next-token loss keeps going down out to ~10M tokens on their own repos. - RULER benchmark at 10M: 100% single-needle, 90% multi-key retrieval, 45% multi-hop QA. - Agentic, run by Mirendil with OpenCode, 1 attempt: 73.4% SWE-bench Verified, 53.2% SWE Atlas QnA, 44.2% DeepSWE, 6.1% Terminal-Bench 4.0. Larger versions are still in training: Base-Large (0.5–2T) is 80%+ through pretraining, XL (2–10T) 20%+.