OpenAI is offering $1M to build an AI that fits in 16 MB

OpenAI just launched Parameter Golf — an open competition where anyone can try to build the smartest possible AI language model that fits inside 16 megabytes. That's smaller than most phone photos. The prize pool: $1 million in computing credits. And yes, it's also a job interview.

The rules: tiny model, big brain

The challenge is deceptively simple. Participants must build an AI language model where the entire package — the model itself plus all training code — weighs under 16 MB (exactly 16,000,000 bytes). For context, ChatGPT's underlying models are estimated to be thousands of times larger.

The models are evaluated on a dataset called FineWeb using a metric called "bits-per-byte" — essentially, how efficiently the model can predict what comes next in a text. It's tokenizer-agnostic (meaning it doesn't matter how you break words into pieces), so participants can get creative with their approaches.

Why this matters beyond the competition

Right now, the AI industry has a size problem. The most capable models require massive data centers and expensive hardware to run. If researchers can figure out how to make tiny models punch above their weight, it opens the door to:

It's also a hiring pipeline

OpenAI isn't hiding it: they're using this competition to find talent. The company plans to hire a small group of junior researchers in June, and top Parameter Golf performers will be invited for interviews. The target: students, competition winners (like math and science Olympiad participants), and self-taught researchers.

This comes as the AI talent war intensifies. Meta has reportedly offered compensation packages up to $300 million to poach top AI researchers — making creative recruitment strategies like this essential for staying competitive.

The current leaderboard

So far, only the baseline model sits on the public leaderboard — a 9-layer model with a score of 1.2244. A longer "unlimited compute" baseline scored 1.2074, showing there's significant room for optimization even within the 10-minute training window.

The GitHub repository (241 stars and climbing) includes starter code, evaluation scripts, and instructions. You can train locally on an Apple Silicon Mac or rent cloud GPUs.

How to participate

# Clone the repo
git clone https://github.com/openai/parameter-golf.git
cd parameter-golf

# Install dependencies
pip install -r requirements.txt

# Start training locally
python train.py

To submit for the official leaderboard, you'll need to open a pull request with your model, a README explaining your approach, and reproducible training logs. The competition runs until April 30, 2026.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News