Mirendil, an AI lab founded a little over a year ago by former Anthropic researchers, has unveiled Fern 0.1, its first model. The headline is the context window: up to 10 million tokens, the amount of text a model can consider at once. According to the company, that is about 23 MB of code, nearly half a million lines, or days of an agent's work held in full. For comparison, the most widely used models today reach about one million tokens.
Fern was pretrained from scratch on Mirendil-Base, a 0.1-0.5 billion parameter model, and then received a short round of post-training. Mirendil presents it as its most efficient model, built for high-volume agentic tasks: auditing the training data of other models, debugging a training run from its logs, and reading agent traces end to end. It accepts text and code as input and is available only through early access requests, with no date for general availability.
On results, the company reports a basic test on the RULER benchmark, which checks how well a model actually uses long context. The most cited figure, however, comes from an internal test: at the maximum context length, Fern answers correctly 45% of the questions that require linking multiple pieces of information. Mirendil promises more comprehensive results in the coming weeks.
The company also stresses how the model was built: according to Mirendil, one researcher working with agents carried out the process, with Mirendil-Base pretrained in 3.5 months for under $10 million in compute, and agents writing over 99% of the code. These are figures the company reports and that no third party has verified.
Fern is also the first of three models. Two larger versions, Mirendil-Base-Large at 0.5-2 billion parameters and Mirendil-Base-XL at 2-10 billion, are being pretrained. Their sizes come only from the company, and no release dates have been set.
The startup is backed by the venture firm a16z, which led its seed round in June 2026. Meanwhile the long-context market is crowded: in recent weeks AlexTech has covered a string of releases with one-million-token windows, from Aleph Alpha to Mercury 3. Fern aims at a figure ten times larger, but the real test will be its performance on long tasks, not the number on the spec sheet.
No comments yet. Be the first!