# Mirendil raises $1 billion for AI research tools > The startup is developing a model and agent environment for AI researchers. By BIGCHANGE.AI Editorial Published: 2026-11-12T19:46:38.143Z Updated: 2026-11-12T19:46:38.143Z Canonical: https://bigchange.ai/blog/mirendil-billion-ai-research-tools ![Illustration of a researcher at a computing console.](https://bigchange.ai/api/media/file/research-console-v3.png) AI-generated illustration for BIG CHANGE. Mirendil has raised $1 billion in a Series A round led by Kleiner Perkins and Clearlake Capital, the AI research startup announced on November 12. The money backs its ambition to build tools that help smaller teams carry out work normally associated with large AI laboratories. ## The big change - Mirendil is pursuing both models and an agent system for research and engineering. Its financing supports that effort, but the company’s account of what its systems can do is not an independent assessment of their reliability. - Fern is a model intended to work with long stretches of code and text. A large input limit does not, by itself, show that a model can understand every detail or make reliable decisions across that input. - Lor is an agent environment in private beta. Access and evaluation matter to researchers: an invitation to test a tool is different from an established, generally available service with independently demonstrated results. ## A model for long research tasks In October, Mirendil described training [Mirendil-Base](https://mirendil.com/news/pretraining-at-scale-without-a-pretraining-team/) with one researcher working alongside AI agents. It says agents wrote most of the code and ran experiments, while the researcher decided what to try and which findings to trust. The company reported 3.5 months and under $10 million in compute, including experiments, data processing and agent token spending. These are company-reported conditions, not an independently audited comparison with a conventional team. Pretraining teaches a model patterns from large amounts of data before it is adapted for particular uses. Mirendil says it used publicly available material and did not buy or license additional data for this experiment. Public availability alone does not settle every question a potential user may have about provenance, rights or suitability for a particular application. The company’s subsequent [Fern announcement](https://mirendil.com/news/introducing-fern/) describes Fern 0.1 as Mirendil-Base after a short additional training stage, with a context window of up to 10 million tokens. A context window is the amount of input a model can consider in one request. The intended work includes reading training logs or long agent traces, where a relevant clue might be far from the beginning of the record. Mirendil reports tests on the underlying base model using RULER, a long-context benchmark. At 10 million tokens it lists 100% for a single-needle retrieval task and 51% for multi-hop question answering, with 100 examples per task. Scores used exact matching, with thinking disabled. Those are different tasks and measures, not one overall accuracy rate; the results are the company’s, not an independent replication. [RULER’s authors](https://arxiv.org/abs/2404.06654) explain why that distinction matters: finding a planted fact in a long input tests a narrower skill than tracing or combining information. Their benchmark adds tasks beyond simple retrieval, but its authors also caution that synthetic tests cannot replace realistic workloads. ## Agents and access In a separate October [description of Lor](https://mirendil.com/news/meet-lor/), Mirendil says an agent can divide a job among other agents, run commands on machines the user has added and request compute for experiments. The company gives the example of building test environments from PyTorch pull requests. That is a company-reported run, not evidence that all generated environments or fixes are correct. The division of labor matters. An agent might prepare several experiments in parallel, but a researcher must still decide whether the runs test the intended hypothesis and whether their results justify a conclusion. Access to machines and the ability to execute commands also make the choice of permitted environments and review of changes relevant to a trial. Lor is an invite-only, free beta for macOS, Linux and the web, and lists Fern among its models. Fern’s standalone page still says access is coming soon and invites early-access requests. Neither page establishes a public model-weight release. The [funding announcement](https://finance.yahoo.com/technology/ai/articles/mirendil-raises-1b-democratize-frontier-173000747.html), distributed by PR Newswire, describes a larger ambition to make frontier AI research more accessible. For researchers assessing that promise, the practical next steps are narrower: obtain access, establish the terms for data and compute, and compare the system’s work with results they can reproduce and check. ## Sources - [Mirendil funding announcement](https://finance.yahoo.com/technology/ai/articles/mirendil-raises-1b-democratize-frontier-173000747.html) — Company release distributed by PR Newswire. - [Mirendil-Base research](https://mirendil.com/news/pretraining-at-scale-without-a-pretraining-team/) — Company description of its training experiment. - [Introducing Fern](https://mirendil.com/news/introducing-fern/) — Model description, reported tests and access. - [Meet Lor](https://mirendil.com/news/meet-lor/) — Company product description and beta terms. - [RULER benchmark research](https://arxiv.org/abs/2404.06654) — Benchmark design and limitations.