Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context
Mirendil has released Fern, its first AI model. Mirendil Fern is a from-scratch coding model with a context window of up to 10M tokens. According to Mirendil, that equates to roughly 23 MB of code, or around half a million lines—a codebase on the scale of Megatron-LM held in context at once.
Fern 0.1 is the first version, built from Mirendil-Base after a short round of post-training. For developers, this means you can load an entire repository, a full training log, or days of agent traces in a single context window.
TL;DR
- Size: 0.1 to 0.5T parameters (exact count and active parameters not disclosed); context up to 10M tokens.
- Runs on: Hosted early access only; no open weights or quantized builds announced.
- Performance: Strong long-context retrieval, mid-pack agentic coding versus open-weight models, well behind frontier models.
- Best: 100% on RULER single-needle retrieval at every length from 32K to 10M tokens (base model).
- Worst: 6.1% on Terminal Bench 4.0, versus 66.4% for Claude Opus 5.5 in Mirendil’s own comparison table.
What Is Mirendil Fern?
Fern is a language model from Mirendil, an AI R&D lab. It is built on the company’s Mirendil-Base model and refined through a short post-training round. The headline capability is an unusually large context window of up to 10M tokens, aimed at workloads where developers need to reason over very large codebases, lengthy training logs, or extended agent traces without chunking or retrieval pipelines.
As of its 2026 launch, Fern 0.1 is available through hosted early access only. Mirendil has not announced open weights or quantized builds, which limits local deployment for now. In its positioning, Fern sits in a crowded field of long-context and agentic coding models, trading some agentic benchmark performance for extreme context length.
Performance Snapshot
Mirendil’s own benchmarks show a clear specialization:
- Long-context retrieval: Fern’s base model scores 100% on RULER single-needle retrieval at every tested length from 32K to 10M tokens.
- Agentic coding: On Terminal Bench 4.0, Fern 0.1 scores 6.1%, compared with 66.4% for Claude Opus 5.5 in Mirendil’s table—an indication that the model is not yet competitive with frontier models on terminal-based agentic coding tasks.
Taken together, the results suggest Fern 0.1 is best suited to tasks that depend on ingesting and retrieving from extremely long inputs, rather than autonomous multi-step coding in terminal environments.
What This Means for Developers
For teams working with large monorepos, lengthy logs, or multi-day agent traces, a 10M-token context window can simplify workflows that would otherwise require retrieval-augmented generation (RAG) or hierarchical chunking. However, because Fern 0.1 is hosted early access only, adoption will depend on API availability, pricing, and latency—details Mirendil has not yet fully disclosed.
The 2026 model landscape increasingly emphasizes long-context capabilities, with several labs pushing context windows into the millions of tokens. Fern 0.1 is Mirendil’s opening entry in that race, and its early benchmark profile suggests a focused tool rather than a general-purpose frontier model.
via MarkTechPost