Researchers from Tsinghua-affiliated labs and collaborators released OpenMLE, a full open-source stack for studying "AI4AI" — AI systems that get better at the process of building AI — and, on top of it, trained and released the weights for Frontis-MA1, a 35-billion-parameter model specialized for machine-learning engineering tasks. Rather than treating model-writes-code-once as the target, the team built the model around four repeatable operators (Draft, Improve, Debug, Crossover) that mirror how a human ML engineer actually iterates: write a first attempt, improve on it, fix what's broken, and recombine ideas from different attempts. These operators are trained with execution-grounded supervised fine-tuning and reinforcement learning, then chained together into a long-horizon search process. The headline result is a jump from a 39.39% medal average on MLE-Bench Lite for the base model up to 71.21% with the full search pipeline, done under a constrained 12-hour budget on a single consumer-grade RTX 4090 — a result the authors report exceeds a GPT-5.5-plus-Codex baseline and approaches larger closed systems. The reason this matters to developers isn't just the benchmark number: it's that the whole stack, including the training gym, RL code, and search framework, is released publicly, so teams can reproduce or adapt the recipe rather than trust a vendor's claims. For anyone building agents that do real ML engineering work — feature engineering, model selection, hyperparameter search, or general "write code, run it, look at the result, try again" loops — this is a concrete blueprint for how to structure that iterative loop as trainable operators instead of a single monolithic prompt, and evidence that smaller, specialized, openly-released models can meaningfully close the gap with much larger frontier systems on well-scoped engineering tasks when the training process is built around the actual shape of the work.