Parameter Golf draws 1,000+ entrants in eight‑week ML challenge, spotlights coding agents' role

The Parameter Golf challenge, an eight‑week open machine‑learning contest, attracted more than 1,000 participants and generated over 2,000 submissions. The event tested participants on compact, fast‑training models under strict resource limits, offering a focused benchmark for efficiency‑oriented research. Competitors were tasked with minimizing held‑out loss measured on a fixed FineWeb dataset while adhering to two hard constraints: model weights and training code had to fit within a 16 MB artifact limit, and each training run was capped at 10 minutes on an 8×H100 training budget. Those limits framed the contest as an exercise in extreme model compression and rapid convergence rather than raw scale.

The artifact and time caps forced entrants to prioritize parameter efficiency, streamlined training pipelines, and clever engineering trade‑offs. With only 16 MB for weights and code, common approaches that rely on large checkpoints or heavyweight training logic were ruled out, pushing teams toward compact architectures, aggressive quantization, parameter sharing or other size‑saving techniques; the 10‑minute training window further favored methods that converge quickly or exploit efficient optimizer and initialization choices. Organizers reported substantial engagement: over 1,000 participants submitted more than 2,000 entries across the competition period. That volume indicates active iteration and experimentation under the contest constraints, with many teams likely exploring multiple architectures and training strategies to balance the size, speed and held‑out loss objectives.

The challenge also underscored the growing role of AI coding agents in model development. Participants leaned on agent tools to accelerate code generation, automate tuning loops and manage repetitive engineering tasks, making agent‑assisted workflows a notable feature of the event’s experimental landscape. To help entrants get started and ensure reproducible evaluation, baseline code and evaluation scripts were published on GitHub. The shared repository provided a common starting point and a consistent scoring pipeline so submissions could be compared fairly under the contest’s artifact and time constraints.