OpenAI has launched Parameter Golf, an open research competition that challenges anyone to squeeze an AI model into just 16MB while training it in under 10 minutes. The prize pool is $1 million in computing credits through partner RunPod, and the competition is open to anyone 18 or older in supported countries.

What Happened

Parameter Golf flips the usual AI scaling race on its head. Instead of building bigger models, participants must build the smallest possible model that still performs well on compression tasks. The rules are strict: combined model weights and training code must stay under 16MB, and training is limited to just 10 minutes on 8 NVIDIA H100 GPUs.

OpenAI has published a full announcement with competition details, alongside a GitHub repository containing baseline models, evaluation scripts, and a public leaderboard so participants can track their progress against each other in real time.

Models are evaluated on compression performance against a fixed FineWeb dataset. The deadline to submit is April 30, 2026.

Why It Matters

The AI industry has spent years in a parameters arms race, with models growing from billions to trillions of parameters. Parameter Golf pushes in the opposite direction, toward efficiency. Smaller, faster models are critical for running AI on phones, edge devices, and resource-constrained environments where cloud access is not guaranteed.

Compression research also has implications for cost. If the same quality can be achieved with dramatically fewer parameters, inference costs drop accordingly. That matters for creators who pay per token or per image generation.

The competition format itself signals a shift in how OpenAI thinks about open research. By providing baseline code, public leaderboards, and generous computing credits, they are lowering the barrier for independent researchers and small teams to contribute meaningful work.

Key Details

  • Prize: $1M in computing credits via RunPod
  • Constraints: 16MB total for weights and training code, 10 minutes on 8 H100 GPUs
  • Evaluation: Compression performance on a fixed FineWeb dataset
  • Eligibility: Anyone 18+ in supported countries
  • Deadline: April 30, 2026
  • Bonus: Top performers may receive invitations for OpenAI job interviews

What to Do Next

If you have experience with model compression, distillation, or quantization techniques, this competition is worth a serious look. The $1M prize pool and potential recruiting pipeline make it one of the more attractive open research competitions running right now.

Even if you are not competing, the public leaderboard and open-source submissions will be a valuable resource for understanding the state of the art in AI compression. Techniques that emerge from this competition could shape how creative AI tools run on local hardware in the near future, a trend already gaining momentum with the recent wave of open-source creative AI models.