Mirendil releases Fern 0.1, its first model, with a 10M-token context window

Mirendil has released Fern 0.1, its first model, with a 10M-token context window. According to a MarkTechPost report published October 27, 2026, the model was pretrained from scratch and is aimed at agentic AI R&D work.

Key facts

• Model: Fern 0.1
• Developer: Mirendil (first model release)
• Context window: 10M tokens
• Training: pretrained from scratch
• Target use: agentic AI R&D work
• Long-context result: 100% on RULER single needle at 10M
• Agentic coding: trails frontier models
• Pricing: not yet disclosed
• Parameter count: not yet disclosed
• Training cutoff: not yet disclosed

Long-context performance

MarkTechPost reports that Fern 0.1 scores 100% on the RULER single needle test at the full 10M-token context length. Single needle retrieval tests whether a model can locate one planted fact inside a very long input.

Where it falls short

The report states that Fern 0.1 trails frontier models on agentic coding. That places the model's strengths in long-context work rather than in coding-agent tasks, where the report says it is behind the leading systems.

What this means

Fern 0.1 pairs a 10M-token window with a perfect single needle retrieval score at that length, and Mirendil is positioning it for agentic AI R&D work. The reported gap on agentic coding is a concrete limit for anyone considering it for coding agents. Because the version is numbered 0.1 and this is Mirendil's first model, evaluation on workloads beyond single needle retrieval will determine how far the long-context result carries over to real tasks.