Mirendil unveils Fern 0.1: a 10 million-token context window, but a weak agent

Mirendil, a startup founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, has announced its first model, Fern 0.1. Its main feature is a 10 million-token context window. A token is a small piece of text, roughly a word or part of one. For code, 10 million tokens comes to about 23 MB, or roughly half a million lines that the model sees at once. By comparison, the other models Mirendil benchmarked against were measured only up to about 1 million tokens. But on coding quality and agent performance, Fern still trails the leaders by a wide margin.

Who needs 10 million tokens

Mirendil is betting not on ordinary programmers but on teams that train neural networks themselves. These people have to work through huge volumes of technical data, and a typical window of a few hundred thousand tokens is too cramped for that. The company suggests loading the following into Fern in full:

• training logs, meaning the records in which a model writes down losses, failures, and metrics day after day;
• records of AI agents' work over several days;
• chunks of training data, to find junk and duplicates in them.

Mirendil also measured how much long context actually helps. It tested how well the model predicts the next fragment of code. Fern's predictions improve as it reads through a repository: the more code it has seen, the more accurately it guesses the continuation. Of all the models compared, only Fern reaches 10 million tokens; the others were measured only up to about 1 million.

There is a caveat. At the same length of 1 million tokens, Kimi-K3 and Beam predict code more accurately than Fern. Both models have already been through post-training, while Mirendil is comparing them against its base version, meaning the model after initial training on text. The company itself notes that post-training usually degrades this metric slightly. So the comparison doesn't favor Fern, but it isn't entirely fair to either side.

Where Fern still loses

On tasks where the model has to act as an agent, the results are weak. Terminal Bench 4.0 tests whether an AI can handle command-line tasks, where it has to type commands itself and make sense of their output. Fern 0.1 scored 6.1% there. Other models' results:

• Claude Opus 5.5: 66.4%;
• GPT-6 Astra: 57.9%;
• GLM-5.3: 41.9%.

The picture is similar on DeepSWE v1.1 and SWE-bench Verified. These are sets of real tasks involving edits to code in existing projects: the model has to fix a bug or extend a feature. According to Mirendil, Fern lags frontier models by about half on DeepSWE and by a quarter on SWE-bench Verified. It does beat some open models.

Its best result is on SWE-Bench Pro v2-Hard, a harder variant of such a set: 76.5%. That matches Beam and is still below frontier models.

Mirendil's explanation is simple: Fern 0.1 is a base model after a short training run. The company promises to release new versions often.

Size and plans: two larger models

The company is vague about the scale of Fern 0.1: between 100 and 500 billion parameters. That is a fivefold spread, so this figure says little about what it would really cost to run the model.

The plans are noticeably more ambitious. Mirendil is already training two more models:

• the first with 500 billion to 2 trillion parameters;
• the second with 2–10 trillion parameters.

Access to an early version of Fern is available on request: you need to contact the startup.

What this means for engineers

Fern 0.1 won't replace Claude or GPT as a coding agent for now: 6.1% on Terminal Bench speaks for itself. If your task is to read a huge training log or check a large chunk of a dataset, 10 million tokens may be more useful than a few extra points on agent benchmarks. But the claimed advantage is so far backed only by Mirendil's own comparisons, and not under the most favorable conditions for Fern. The real test begins when the model reaches independent teams and larger versions ship. Then it will become clear whether it can work with long context as well as it can hold it.