Mirendil Raises $1 Billion for AI Training Automation and Model Development
- Cargo
- Machine Learning
- Shipped
- Shipper
- Maria Deutscher
- Readers
- 323
Mirendil’s $1 billion Series A backs Lor, an agent platform for automating model training, and further development of its coding models.

Mirendil plans to apply a $1 billion Series A to software that automates parts of AI model development. Its platform, Lor, assigns training and evaluation work to AI agents, while the company’s own coding models offer an early example of what it says this approach can produce. The funding was jointly led by Kleiner Perkins and Clearlake Capital, with Google and more than half a dozen other participants.
Automating the model development loop
Training a large language model involves repeated experiments. Researchers adjust an algorithm, run training and evaluation, inspect the results, then decide what to change next. Small choices in model architecture, data preparation and training methods can affect output quality, so teams often need to test many variants before settling on an approach.
Mirendil says Lor can coordinate much of that process through hundreds of AI agents. The company describes the platform as handling tasks across a training run, including infrastructure setup, algorithm selection, data collection and model refinement. It also says Lor can identify and fix errors that arise during evaluation. The practical claim is that one researcher could direct work that would usually involve several specialists, though the account of the company’s project does not provide a benchmark comparing Lor’s results or productivity against a conventional team.
That distinction matters for assessing the software. Automating code generation and routine experiment steps could reduce the amount of hands-on work required to explore model variants. But the quality of the resulting model still depends on the objectives, evaluation methods and data supplied to the process. The reported project offers a case study of Lor in use; it does not establish how reliably the platform performs across different model types or research tasks.
A coding model built with agent assistance
Earlier this year, Mirendil assigned one researcher to develop a coding model intended to compete with leading open-source LLMs. The project took about 3.5 months and cost $10 million, according to the company. Mirendil says its agents wrote more than 99% of the code powering Mirendil-Base, the model produced by the effort.
The agents also set up infrastructure, selected a training algorithm and gathered web data. They then refined the dataset and the model’s architecture to improve output quality. Those details show the breadth of work Mirendil says it has automated, but the available account gives no independent evaluation scores, comparison results or breakdown of the $10 million cost. Readers therefore have a clear description of the process and its reported scale, but limited evidence for judging how Mirendil-Base performs against its intended competitors.
Mirendil-Base underpins Fern, a specialized coding model the company unveiled last month. Fern is designed for programming challenges that AI labs commonly encounter. Mirendil says it can accept 10 million input tokens in a prompt—ten times the stated context length of Claude Opus 5.5—and equates that input capacity to roughly half a million lines of code. A large context window can let a model take in more code at once, but token capacity alone does not show how accurately it can reason over that material or complete coding tasks.
Scaling plans and open questions
Fern uses a version of Mirendil-Base with between 100 billion and 500 billion parameters. Mirendil has also previewed two larger versions of Mirendil-Base, ranging from 500 billion to 10 trillion parameters, and says it intends to use them to build more capable LLMs. These parameter ranges indicate the scale the company is pursuing; they do not, by themselves, establish model quality or the compute and serving costs required to train and deploy the systems.
The company was launched last year by Behnam Neyshabur and Harsh Mehta, both of whom previously worked at Google DeepMind and Anthropic. Neyshabur co-invented SAM, an algorithm intended to improve model performance on tasks beyond those used in training. Mehta worked on pretraining projects, the initial stage in an AI model’s development. Their backgrounds align with Mirendil’s focus on automating training work and building models through that process.
The $1 billion investment gives Mirendil substantial backing for its model development plans, while Lor’s central proposition is that agent coordination can compress work usually spread across a larger research team. The next meaningful evidence will be reproducible performance results for Mirendil’s models and clearer measures of how much time, compute and human oversight Lor requires across projects. Until then, the reported coding-model effort is a notable demonstration of the company’s workflow, with its general effectiveness still to be established.