Mirendil has released Fern 0.1, its first AI model, which is a from-scratch coding model with a context window of up to 10 million tokens.
Key points
- Mirendil released Fern 0.1, its first AI model.
- Fern 0.1 features a 10 million token context window.
- The model is designed for AI R&D and coding tasks.
- It achieved 100% on RULER retrieval up to 10M tokens.
- Fern 0.1 is available via hosted early access only.
This context window is equivalent to approximately 23 MB of code, or about half a million lines, allowing it to handle codebases on the scale of Megatron-LM.
Fern 0.1 is the initial version, built upon Mirendil-Base after a brief post-training period, and is intended for developers to load entire repositories, training logs, or agent traces at once.
By the numbers
- Context window size
- 10M tokens
- Terminal Bench 4.0 score
- 6.1%
The model is aimed at AI R&D work, with target uses including working across larger codebases, agentic coding, research code analysis, debugging long logs, and auditing pretraining shards.
Fern is built on Mirendil-Base, which a researcher pretrained in 3.5 months with under $10 million in compute, including experiments and data processing.
Over 99% of the code for Fern was written by AI agents, with a researcher guiding the process. Training data was sourced exclusively from public repositories like GitHub, public code, web pages, and public datasets, with no distillation, synthetic, or licensed data used.
The model was coded from scratch using PyTorch and JAX. Mirendil Fern's performance includes 100% accuracy on RULER single-needle retrieval across all tested lengths up to 10 million tokens, but it scored 6.1% on Terminal Bench 4.0, compared to Claude Opus 5.5's 66.4% in Mirendil's own comparison.
Fern is currently available via hosted early access only, with no open weights or quantized builds announced.
Sources and references
Original report
Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Linkmirendil.com/news/introducing-fern (opens in a new tab)
- Linkmirendil.com (opens in a new tab)
- Linkmirendil.com/news/pretraining-at-scale-without-a-pre… (opens in a new tab)
- Paperarxiv.org/abs/2404.06654 (opens in a new tab)
- Modelhuggingface.co/moonshotai/Kimi-K3 (opens in a new tab)
- Linkbuild.nvidia.com/z-ai/glm-5-3 (opens in a new tab)
Ref 00N64ZC8
Cite this
APA
Interestana AI Editorial. (2026, October 27). Mirendil Releases Fern 0.1 AI Coding Model With 10M-Token Context. Interestana. https://interestana.com/articles/mirendil-launches-fern-01-its-first-model-with-10m-token-context-00n64zc8
MLA
“Mirendil Releases Fern 0.1 AI Coding Model With 10M-Token Context.” Interestana, 27 Oct. 2026, interestana.com/articles/mirendil-launches-fern-01-its-first-model-with-10m-token-context-00n64zc8.
Chicago
Interestana AI Editorial. “Mirendil Releases Fern 0.1 AI Coding Model With 10M-Token Context.” Interestana, October 27, 2026. https://interestana.com/articles/mirendil-launches-fern-01-its-first-model-with-10m-token-context-00n64zc8.
BibTeX
@misc{interestana_00n64zc8,
author = {{Interestana AI Editorial}},
title = {{Mirendil Releases Fern 0.1 AI Coding Model With 10M-Token Context}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/mirendil-launches-fern-01-its-first-model-with-10m-token-context-00n64zc8},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/27/mirendil-fern-10m-token-context-model/}
}Credit the source
Original reporting: MarkTechPost, “Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context”, https://www.marktechpost.com/2026/10/27/mirendil-fern-10m-token-context-model/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe