Mirendil Trains 100B-Parameter Coding Model With One Researcher and AI Agents

On October 23, Silicon Valley startup Mirendil announced that it has trained Mirendil-Base, a coding-specialized foundation model with hundreds of billions of parameters, using just one researcher and AI agents over three and a half months, at a cost of under $10 million. The company said more than 99% of the code was written by agents, with the human researcher responsible for setting direction, making decisions, and intervening at key milestones.

Mirendil said the training did not rely on existing pretraining frameworks or inherited code. Training infrastructure, model architecture, training algorithms, and data pipelines were built from scratch by the researcher and agents. All training data came from public sources such as GitHub repositories and public web pages; the company said it purchased no data, performed no distillation, and used no synthetic data.

In a cross-entropy benchmark on a new code repository, Mirendil-Base scored 0.222, trailing only Kimi K3 at 0.193 and ahead of GLM-5.3, DeepSeek-V4, and Nemotron. The company also revealed that the larger Mirendil-Base-Large is 80% trained, while Mirendil-Base-XL is 19% trained.

Founded in December 2025 by two former Anthropic researchers, Mirendil secured a $200 million seed round in June at a $1 billion valuation, with investors including a16z, Kleiner Perkins, and Nvidia. The company aims to help scientists build their own AI models rather than merely use AI as a research aid.