Mirendil’s First Model Fern Scores 100% on 10M-token Long Context Retrieval, Lags Leading Models in Agent Coding
Mirendil has released Fern 0.1, its first large language model trained from scratch. The model boasts a 10-million-token ultra-long context window targeting agent AI development scenarios. It earned a perfect score in RULER’s 10M-token single-needle retrieval test but underperforms frontier models in agent coding, taking a differentiated niche approach.
On October 27, 2026, AI company Mirendil officially launched Fern 0.1, its first large language model. Trained from scratch, the model features a 10-million-token context window and is primarily designed for agent AI development use cases.
For readers unfamiliar with technical jargon, a context window refers to the maximum amount of text a model can process in a single pass. Tokens are the basic unit of text processing for large models; 10 million tokens roughly equal about 15 million Chinese characters, enough to accommodate all code, supporting documentation, and months of team communication logs for a mid-sized project.
Long context has been a core focus in the LLM industry over the past two years, but specs on paper don’t translate to real-world performance. The longer the text, the higher the risk a model will miss critical details in the middle—a widespread industry challenge. Fern 0.1 scored 100% on RULER’s authoritative long-context benchmark test for 10-million-token single-needle retrieval. This task hides a specific piece of information in a random position within a large corpus, requiring the model to extract it precisely; a perfect score means it retains key information even at the 10-million-token length.
Fern 0.1’s current capabilities are imbalanced. In practical tasks like agent coding, it lags behind the top-tier leading models. The company has no plans to build a general-purpose LLM, with all optimizations focused on the needs of development-oriented agent models.
As long context parameters grow increasingly competitive, hitting the 10-million-token mark is no longer a novelty. The real bottleneck has always been the ability to retain information accurately, not just token count. Fern 0.1 takes a niche approach, skipping the 'jack-of-all-trades' high-score strategy to prioritize full accuracy for long contexts. Whether this product strategy will resonate depends on if development users will accept its coding limitations in exchange for long-text stability.
Source: MarkTechPost, author Michal Sutter
发布时间: 2026-10-27 20:57