Frontis releases Frontis-MA1, a 35B parameter AI4AI model designed for recursive self-improvement in machine learning engineering, alongside the open-source OpenMLE stack. The system integrates verifiable task environments, operator learning, and long-horizon search to enable an agent that improves its own code through execution-grounded training.

  • Frontis-MA1 is post-trained using four atomic program-evolution operators: Draft, Improve, Debug, and Crossover.
  • On MLE-Bench Lite with a 12-hour budget on one RTX 4090, the model improves Medal Average from 39.39% to 60.61% using OpenMLE-Evo.
  • With OpenMLE-Evo-Max, Frontis-MA1 reaches 71.21%, exceeding GPT-5.5 + Codex and approaching GPT-5.6 Sol and Kimi K3.
  • The framework demonstrates transferability on NatureBench Lite, raising Match-SOTA from 50% to 70% when swapping in the trained model.

The release of both the model weights and the full OpenMLE stack aims to enable reproducible research on executable AI4AI systems toward recursive self-improvement.