Mirendil Secures $1B Series A to Automate AI Training and Scale Coding LLMs
Mirendil has raised $1 billion in Series A funding led by Kleiner Perkins and Clearlake Capital to automate AI model training using hundreds of agentic workers.
Artificial intelligence startup Mirendil Inc. has secured $1 billion in early-stage funding to accelerate the development and training of large-scale AI models. The Series A round was co-led by venture firms Kleiner Perkins and Clearlake Capital, alongside contributions from Google LLC and more than six other investors. Mirendil was established last year by former researchers from Google DeepMind and Anthropic PBC. Chief Executive Officer Behnam Neyshabur previously co-invented SAM, a popular algorithm designed to help AI models perform better on tasks outside their initial training data, while Chief Technology Officer Harsh Mehta focused on pretraining workflows during his tenures at DeepMind and Anthropic.
Automating model training with Lor
Mirendil’s flagship offering is Lor, an automated software platform designed to streamline pretraining—the resource-intensive initial phase of constructing artificial intelligence models. Rather than relying on large teams of engineers to handle manual experimentation, Lor distributes model development tasks across hundreds of specialized AI agents.
Building large language models typically requires extensive trial-and-error, as researchers tweak algorithms in minor ways and benchmark competing variants to determine optimal performance. Lor automates this iterative comparison workflow and can independently resolve errors encountered during model evaluations.
To demonstrate the capability of its software, Mirendil tasked a single researcher with creating a competitive open-source coding model. Utilizing Lor’s automation technology, the single developer completed the project in approximately 3.5 months at a total cost of $10 million, without receiving assistance from colleagues. Autonomous AI agents wrote more than 99% of the code for the resulting model, named Mirendil-Base. The agents managed infrastructure setup, selected the training algorithm, gathered web data, and systematically refined both the training dataset and the underlying architecture to enhance output quality.
Scaling context windows and parameter sizes
Using Mirendil-Base as a foundation, the company unveiled a specialized coding model called Fern last month. Fern is tailored to automate complex programming problems frequently faced by artificial intelligence research laboratories.
A key distinction of Fern is its large context processing capacity. The model can handle up to 10 million input tokens in a single prompt, which represents a tenfold increase over Claude Opus 5.5. This capacity enables Fern to ingest approximately half a million lines of code in one request. Fern’s underlying structure is based on a version of Mirendil-Base containing between 100 billion and 500 billion parameters.
Mirendil is also preparing to scale its technology further. Late last month, the startup previewed two enhanced iterations of Mirendil-Base spanning parameter counts from 500 billion up to 10 trillion. The company intends to leverage these massive algorithms to build even more capable future language models.
What it means for developers
Mirendil's approach signals a potential shift in how frontier AI models are engineered and scaled. By substituting manual engineering labor with multi-agent orchestration, the company demonstrated that single-developer teams can execute model builds that previously required massive groups of specialists. For AI research groups and software engineers, platforms like Lor could dramatically lower the human overhead needed to curate training data, configure infrastructure, and fine-tune model architectures.
Furthermore, models with extreme context windows—such as Fern's 10-million-token input capacity—could transform how developers interact with large codebases, enabling comprehensive repository analysis within a single query. As model architectures grow and context limits expand, staying flexible across different model providers becomes increasingly vital for engineering teams. Developers interested in experimenting with diverse model families can try top AI models cheaply through one API at https://apixoai.online, streamlining integration while managing token costs.
If Mirendil successfully scales its 10-trillion-parameter architectures, the resulting models may set new performance benchmarks for automated software engineering and pretraining efficiency.
Source: Mirendil raises $1B in funding to speed up AI model development — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
