Fern 0.1: 10M Token Context from Ex-Anthropic

Startup Mirendil, founded by Anthropic alumni Behnam Neyshabur and Harsh Mehta, introduced the Fern 0.1 model with pretraining from scratch and a context window of up to 10 million tokens: it passes RULER retrieval tests at this length almost perfectly, but currently lags significantly behind competitors in agentic tasks.

Compiled by Sergey KostenchukPublished 2026-10-30Updated 2026-10-30

2026-10-30 Research Anthropic

🤖 Fern 0.1: 10M Token Context from Ex-Anthropic

On October 27, 2026, startup Mirendil, founded by Anthropic alumni Behnam Neyshabur and Harsh Mehta, introduced the first Fern 0.1 model: pretraining from scratch, short post-training, and a window of up to 10 million tokens — approximately 23 MB of code, half a million lines.

🌍 Mirendil is betting on the thesis that “Context is Intelligence”: training logs, agent traces, and entire datasets are placed into the context instead of scaling parameters. Retrieval at 10 million tokens is almost ideal (tracking 100%, multi-key search 91–95%), but multi-hop QA drops to 51%.

👤 Practical interest is currently in the niche of giant homogeneous contexts — entire repositories, logs, dataset cleaning. It is too early to use the model as a coding agent: on Terminal Bench 4.0 — 6.1% compared to 66.4% for Claude Opus 5.5. Early version — available by request via mirendil.com/early-access.

Source 1: https://mirendil.com/news/introducing-fern/ Source 2: https://www.neyshabur.net/

Sources