OpenAI is running a new Model Craft challenge called Parameter Golf focused on building an efficient pretrained model under strict size and compute limits.
The challenge asks participants to minimize held-out loss on a fixed FineWeb dataset split while staying within a 16 MB total artifact limit, covering model weights and training code together, and a 10-minute training budget on 8×H100 GPUs. OpenAI says it provides a GitHub repository with a baseline, the fixed dataset, and evaluation scripts.
How submissions work
According to the source, entrants are expected to fork the repository, improve the model within the stated limits, and submit a pull request with code, logs, a score, and a short write-up. After approval, the result is merged and the leaderboard updates automatically.
OpenAI also states that participants can apply for free Runpod compute credits while supplies last. The Terms and Conditions published on March 25, 2026 confirm the challenge rules and the availability of Runpod credits for competitors.
OpenAI's stated purpose
OpenAI says the challenge is meant to surface researchers and engineers it may want to hire. The page states that standout participants may be invited to interview for job opportunities at OpenAI, and that winning approaches may be featured publicly.