OpenAI launches Parameter Golf to push efficient pretraining under a 16 MB cap

OpenAI said on X on March 18, 2026 that it is opening a new research competition called Parameter Golf. The linked challenge page frames it as an attempt to find the most efficient pretrained model under unusually tight constraints: entrants must minimize held-out loss on a fixed FineWeb dataset while staying inside a 16 MB artifact limit for weights and training code combined, plus a 10-minute training budget on 8×H100s.

That design makes the challenge more than a marketing exercise. OpenAI is forcing participants to optimize for parameter efficiency, training efficiency, and reproducibility at the same time. Instead of rewarding the biggest model or the longest training run, the setup rewards ideas that make small models train better under hard compute and packaging limits. For researchers working on edge deployment, efficient pretraining, or compact foundation models, that makes the contest unusually relevant.

OpenAI is also structuring the process like an open engineering benchmark. The company says it is providing a public GitHub repository with a baseline, a fixed dataset, and evaluation scripts. Participants are expected to fork the repo, improve the model within the size and compute caps, and submit a pull request containing code, logs, a score, and a short write-up. Once a submission is approved, the leaderboard updates automatically.

The company is also using the challenge as a talent funnel. On the challenge page, OpenAI says standout participants may be invited to interview for roles at the company, and that winning approaches may be featured publicly. It is also offering optional Runpod compute support, ranging from quick-start credits to larger grants for advanced competitors, subject to availability and eligibility review.

The broader signal is that frontier labs are paying renewed attention to small, highly efficient models and the engineering discipline around them. Parameter Golf turns that priority into a public contest with measurable rules. If strong entries emerge, the results could surface techniques that matter well beyond the leaderboard, especially for developers trying to squeeze more useful behavior out of smaller model budgets.