Frontis-MA1 Hits 71% MLE-Bench Lite Score on a 4090

Frontis-MA1, a 35B parameter agent, pushed MLE-Bench Lite scores from 39.39% to 71.21% on a single RTX 4090 capped at 12GB VRAM, per the July 30 arXiv paper. That kind of gain used to need a multi-GPU cluster and a research budget most independent labs don't have. Frontis-MA1 got there in a 12-hour per-task window on a desktop graphics card.

The model is post-trained on OpenMLE, an open system built for recursive self-improvement research. It splits the work into three parts: OpenMLE-Gym for verifiable task environments, OpenMLE-RL for operator learning, and OpenMLE-Evo for long-horizon search. Four atomic operators, named Draft, Improve, Debug, and Crossover, get trained through execution-grounded supervised fine-tuning and reinforcement learning. Then they get composed into a search loop that writes and rewrites its own code.

The base setup hit 60.61% medal average; adding OpenMLE-Evo-Max pushed it to 71.21%, beating GPT-5.5 plus Codex on the same benchmark, per the paper. It doesn't clear every bar. Frontis-MA1 approaches but doesn't pass GPT-5.6 Sol or the 2.8 trillion parameter Kimi K3, both much larger models running on far heavier infrastructure.

A related report describes a system called SkillRise: a single policy alternating between solving tasks and rewriting its own skill document. That's the same self-improvement instinct driving Frontis-MA1. Another describes attackers pulling a multi-agent system's inference-time harness out through black-box queries alone. That's the opposite move from a team that openly published its whole stack for free.

The open stack is the bigger story here, not the score. If OpenMLE-Evo gets forked onto a different base model and the medal average doesn't transfer, 71.21% looks like a benchmark artifact, not a capability jump. Nobody outside the paper has tried that yet, and I'd bet against the number holding up cleanly on a second model.