In a new research paper, Hugging Face presents Frontis-MA1, a 35-billion-parameter model designed to push the boundaries of recursive self-improvement (RSI) in machine learning engineering (MLE). The work introduces OpenMLE, an open full-stack system that provides a concrete testbed for AI4AI—AI systems that improve the process of building AI itself.
OpenMLE comprises three key components: OpenMLE-Gym, a verifiable task environment with execution feedback; OpenMLE-RL, for operator learning; and OpenMLE-Evo, for long-horizon search. On this stack, the team post-trained Frontis-MA1 as a meta-evolution agent, aligning training and inference around four atomic program-evolution operators: Draft, Improve, Debug, and Crossover.
The results are striking. On MLE-Bench Lite, under a 12-hour per-task budget on a single RTX 4090 capped at 12 GB VRAM, Frontis-MA1 improved its Medal Average from 39.39% to 60.61% over its base model using OpenMLE-Evo, and reached 71.21% with OpenMLE-Evo-Max—exceeding GPT-5.5 + Codex and approaching GPT-5.6 Sol and the 2.8T Kimi K3.
On held-out NatureBench Lite, both components transferred effectively: with the framework fixed, swapping in the trained model raised Match-SOTA from 50% to 70%; with the model fixed, swapping in OpenMLE-Evo raised it from 20% to 50%. The authors release the model weights and the full OpenMLE stack to enable reproducible research on executable AI4AI toward RSI.