Frontis-MA1 (35B) Reaches 71.21% MLE-Bench Lite Medal Average on a Single 12 GB RTX 4090, Weights Released

arXiv·high signalXBlueskyLinkedInCopy link

Frontis-MA1 (arXiv 2607.28568, July 30) is a 35B meta-evolution agent post-trained on OpenMLE, an open full-stack system for recursive self-improvement research spanning verifiable task environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo). Four atomic program-evolution operators — Draft, Improve, Debug, Crossover — are trained via execution-grounded SFT and RL, then composed into search. Under a 12-hour per-task budget on one RTX 4090 capped at 12 GB VRAM it lifts MLE-Bench Lite Medal Average from 39.39% to 60.61%, and to 71.21% with OpenMLE-Evo-Max, exceeding GPT-5.5 + Codex and approaching GPT-5.6 Sol and the 2.8T Kimi K3; model weights and the full stack are public.