Mirendil introduces Fern, its first model with a 10M token context window

·2026.10.26 09:00

Key point

The model is designed for high-volume agentic tasks, achieving a 76.5% pass rate on SWE-Bench Pro v2-Hard in its initial 0.1 release.

Mirendil releases Fern, its first model, pretrained from scratch with a context window of up to 10M tokens (approximately 23 MB of code).
Fern 0.1, the initial version after short post-training, scores 76.5% on SWE-Bench Pro v2-Hard and 73.4% on SWE-Bench Verified.
In next-token prediction tests, Fern continues to improve cross-entropy up to 23M bytes, whereas comparison models like Kimi-K3 and GLM-5.3 plateau around 1M tokens.
The model is optimized for agentic tasks including debugging training runs, auditing data, and analyzing full agent traces.
Two larger models, Mirendil-Base-Large (0.5–2T params) and Mirendil-Base-XL (2–10T params), are currently in pretraining.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.