Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context

Mirendil, an AI R&D lab, has released Fern 0.1, its first model. It's a coding model built from scratch with a context window of up to 10 million tokens, which the lab says is about 23 MB of code or roughly half a million lines. That's enough to load an entire large repository, a full training log, or days of agent traces at once.

The build story is unusual. Mirendil says the underlying base model was pretrained by one researcher in 3.5 months for under $10 million in total compute. AI agents wrote more than 99 percent of the code, and training used only public data, with no distillation, synthetic data or licensed data. For now, Fern is available only through hosted early access. No weights have been released.

We find the retrieval results impressive. On the RULER benchmark, the base model found a single planted needle 100 percent of the time at every length from 32,000 to 10 million tokens. Its prediction loss on code also kept improving past the point where other models stop. For anyone debugging a huge training run or auditing a large codebase, a model that can actually see all of it is useful.

But Mirendil's own tables show the limits, and we give the lab credit for publishing them. Multi-hop reasoning fell to 45 percent at 10 million tokens, and that figure rests on just 20 examples per cell. On agentic coding, Fern scored 6.1 percent on Terminal Bench 4.0, compared with 66.4 percent for Claude Opus 5.5 in the same table, and it trails every frontier and large open model listed. At equal context, Kimi-K3 was the better predictor. The fair reading is that Fern can hold more but doesn't yet reason over it as well.

The bigger story may be how it was made. One researcher steering agents to build a base model for under $10 million is a striking claim about how cheap and automated frontier-adjacent research is becoming. If that holds up, more small labs will try. That's healthy for competition. It also means less human review of the code behind models that may end up deployed widely, and we'd want to see how Mirendil checked what its agents wrote.

One more caution. A model that reads entire repositories and agent traces is exactly the kind of tool that will get pointed at proprietary code and sensitive logs. Hosted-only access means that data goes to Mirendil's servers, and the license hasn't been disclosed. Teams should get clear answers on data handling before uploading their crown jewels.

Context windows set how much text a model can consider at once. The open-weight models Mirendil compared it against advertise limits of about one million tokens. Mirendil says Fern is the smallest of three planned models, with larger versions partway through pretraining. Benchmark comparisons here come from Mirendil's announcement, using figures from model cards, leaderboards and the lab's own runs.