The company has transitioned to post-training its first large-scale coding model after successfully substituting a full team of engineers with AI agents. These agents authored over 99% of the codebase and managed the research experiments, while 1 human researcher directed the strategy and decided which results to trust. Mirendil plans to release API access and independent evaluations shortly, and the project was completed without using knowledge distillation to reach its performance levels.
Mirendil-Base contains between 100B and 500B parameters and was developed from scratch in 3.5 months using less than $10M in compute. The organization is now scaling this approach to a “Large” variant with 500B to 2T parameters and an “XL” version targeting up to 10T parameters. The initiative aims to democratize frontier AI development by utilizing recursive self-improvement to drastically reduce the expertise and financial capital required to train foundation models.