Mirendil has released Fern 0.1, its first AI model, which can hold a context window of up to 10 million tokens. That capacity translates to roughly 23MB of code or approximately half a million lines. The company states this allows a single session to contain a codebase as large as Megatron-LM, full training logs, or days of agent traces.
The model is built from scratch on Mirendil-Base following a brief post-training phase. It is currently available via hosted early access only, with no open weights or quantised builds announced. While retrieval performance is strong, agentic coding capabilities sit in the middle pack against open-weight models and lag significantly behind frontier systems.
What is Mirendil Fern?
Fern comes from Mirendil, an AI research and development laboratory. The tool is designed for research and development tasks rather than general conversation. Mirendil lists six specific use cases:
- Working across larger codebases
- Agentic coding that executes commands until tests pass
- Research code involving training loops, data pipelines and evaluations
- Debugging training runs from long logs and stack traces
- Reading full agent or training traces from start to finish
- Auditing pretraining shards for duplicates, contamination and junk
How was Fern built?
The model sits on Mirendil-Base, which the company states one researcher pretrained in 3.5 months. Total compute cost was under $10 million, covering experiments, agent token spend and data processing. AI agents wrote over 99% of the code, while the researcher decided what to try and when to intervene.
Training data came only from public sources: GitHub, public code, web pages and public datasets. Mirendil says it used no distillation, no synthetic data and no licensed data. The code was written from scratch using PyTorch and JAX.
Fern is the smallest of three planned models. Mirendil-Base-Large, with 0.5 to 2 trillion parameters, is over 80% through pretraining. Mirendil-Base-XL, with 2 to 10 trillion parameters, is over 20% through.
Does longer context actually help?
Mirendil tested next-token prediction on its own repositories, concatenated into a 42.9MB stream. It measured cross-entropy in bits per byte, where lower is better. Window sizes here are measured in bytes, not tokens.
Fern’s loss fell from 0.246 at a 32K-byte window to 0.155 at 16M and 23M bytes. The other models were scored up to about 1 million tokens, their advertised limit. At a 1M-byte window, Kimi-K3 scored 0.140 and Fern 0.165. So Fern is not the best predictor at equal context. Its advantage is that the curve keeps going past where others stop.
How does Fern score on RULER?
RULER is a standard long-context benchmark. Mirendil ran five RULER tasks on Mirendil-Base, the pre-post-training model. It used greedy decoding, thinking off and exact string matching.
Retrieval holds up well. Multi-hop reasoning drops to 45% at 10M. Each 10M cell uses only 20 examples, versus 100 elsewhere. Mirendil says broader long-context results are coming.
How does Fern 0.1 do on agentic benchmarks?
Mirendil compared Fern 0.1 with open-weight and frontier models. Scores for other models come from model cards, leaderboards or Mirendil’s own runs.
Fern 0.1 beats Inkling and Nemotron 3 Ultra on four of five tests. It ties Beam on SWE-Pro v2 Hard at 76.5. It trails Beam on SWE-Bench Verified, 73.4 versus 81.6. Every frontier and large open model in the table leads it.
How does Fern compare with similar models?
Here is a comparison of Fern 0.1 with Kimi-K3 and GLM-5.3:
- Developer: Mirendil, Moonshot AI, Z.ai
- Total params: 0.1 to 0.5T, 2.8T, 753B
- Active params: Not disclosed, 104B, ~40B
- Context: Up to 10M tokens, 1,048,576 tokens, 1,048,576 tokens
- Input: Text, code; Text, image, video; Text
- License: Not disclosed; Kimi K3 License; GLM-5.3 License
- Weights: Not released; Open; Open
- Access: Early access request; Open weights, APIs; Open weights, APIs
- SWE-Bench Verified: 73.4, 93.4, 95.4
- Bits per byte at 1M bytes: 0.165, 0.140, 0.170
*Figures from Mirendil’s announcement charts.
What it means
Developers can load entire repositories or days of agent traces into a single context window. The model excels at finding specific information within massive datasets, maintaining perfect retrieval scores across the full 10M token range. However, it is not yet a competitor to top-tier models for writing complex software or executing autonomous tasks, scoring poorly on terminal benchmarks compared to established systems.
