OpenAI is trying something different this time—and it’s not about building the biggest AI model.
In a new challenge called Parameter Golf, the company is asking developers to go in the opposite direction: make language models as small and efficient as possible, without losing performance.
The rules are tight. Participants need to build a model that fits within just 16MB and can be trained in under 10 minutes on an NVIDIA 8×H100 GPU setup. The final models will be tested on how well they perform on the FineWeb validation dataset, especially in terms of compression.
What makes this interesting is the shift in thinking. For years, AI progress has largely been driven by scaling—more data, bigger models, higher costs. This challenge flips that idea. Here, the goal is simple but tough: get the best possible results using very limited parameters.
Because of these constraints, developers can’t rely on traditional approaches alone. They’ll likely need to experiment—things like sharing parameters across layers, using low-rank techniques, or even rethinking how text is broken into tokens could make a difference.
To encourage more participation, OpenAI is offering $1 million in compute credits. That support could help smaller teams and independent researchers take part without worrying too much about infrastructure costs.
There’s also a bigger picture behind this. OpenAI is clearly interested in finding people who can solve unusual problems creatively, especially those early in their careers. Challenges like this tend to highlight practical thinking more than just theoretical knowledge.
The competition started on March 18 and will run until April 30. Entries need to follow strict rules, including being fully reproducible and staying within all the given limits.
In many ways, Parameter Golf reflects where AI could be heading next. Instead of just making models bigger, the focus is slowly moving toward making them smarter, faster, and easier to run in real-world conditions.
Also Read: OpenAI Unveils GPT-5.4 Mini and Nano
