OpenAI launches Parameter-Golf: the ultimate challenge to create the most compact language model

Artificial intelligence is reaching a decisive turning point today. While the industry has long focused on giant models with billions of parameters, a new trend is emerging: radical optimization.

OpenAI has just launched the Model Craft Challenge: Parameter Golf, a competition that completely reverses the paradigm. The objective? To train the best language model possible while respecting an extreme constraint of 16 MB. This pioneering challenge could define the standards for future embedded AI, capable of running on smartphones, smart glasses, or low-power devices.

Are you up for a challenge? https://t.co/GNryIDhnut pic.twitter.com/ZX7ZiuhGgu

— OpenAI (@OpenAI) March 18, 2026

The Parameter-Golf Challenge: Pushing the Limits of Compression

The challenge is as simple as it is formidable: create a high-performing language model whose total weight (architecture + training code) does not exceed 16 megabytes, and capable of training in less than 10 minutes on an 8xH100 configuration. Participants are evaluated on their ability to minimize loss on the FineWeb dataset, measured in Bits-Per-Byte (BPB).

This constraint forces developers to explore fascinating technical territories: innovative architectures (Mixture of Experts, parameter tying, recurrence), aggressive compression schemes (extreme quantization, QAT, bitnets), optimized tokenizers, or even unconventional training methods (test-time training, long context). OpenAI explicitly encourages creative submissions, even if they don’t yet break the current record.

Practical Information & Calendar for the Parameter-Golf Challenge

The competition takes place 100% online via GitHub. Participants fork the official repository, develop their solution, then submit a Pull Request with their code, logs, and documentation.

Key Dates

• Opening: March 18, 2026, 10:00 AM PST
• Submission Deadline: April 30, 2026, 5:00 PM PST
• Talent Recruitment: June 2026

Resources Provided

• $1 million in compute credits sponsored by OpenAI and distributed via Runpod
• Pre-downloadable FineWeb dataset with 1024-token vocabulary
• Baseline training scripts (PyTorch and MLX for Apple Silicon)
• Pre-configured Runpod template with a complete environment

Who can participate in the Parameter-Golf challenge?

The challenge is open to all passionate AI developers and researchers, provided they meet these conditions:

• Be a legal resident of a jurisdiction supported by the OpenAI API (list available on the official website)
• Be at least 18 years old and the age of majority in their country of residence
• Not be on any international trade sanctions list

No registration fee is required. Students, Olympic medalists, and recent graduates are particularly encouraged to participate.

Evaluation Criteria & Rules for Parameter-Golf

Submissions are judged on their BPB (Bits-Per-Byte) compression score on the FineWeb validation split. To beat the current leaderboard record, an improvement of at least 0.005 nats is required, with statistical significance p < 0.01 demonstrated over multiple runs.

Strict Rules

• Total artifact ≤ 16 MB (decimal, i.e., exactly 16,000,000 bytes)
• Training time ≤ 10 minutes on 8xH100 for the main leaderboard
• No external access during evaluation (no downloads, no network calls)
• Full reproducibility required, non-reproducible submissions will be disqualified
• Tokenizer modifications are allowed but rigorously scrutinized to avoid calculation biases

OpenAI reserves the right to reject submissions that circumvent the spirit of the challenge.

Resources & Official Links

• Official GitHub Repository: https://github.com/openai/parameter-golf
• Challenge Page: https://openai.com/index/parameter-golf/

Why participate in the OpenAI Parameter-Golf Hackathon?

Beyond the competitive aspect, this challenge offers unique opportunities:

Intensive Learning: You will dive into cutting-edge optimization techniques rarely explored at this scale. Access to compute credits allows for experimentation without budgetary constraints.

Networking and Visibility: The public leaderboard and Discord channels connect you directly with the global community of AI researchers. Innovative approaches, even without breaking records, can be publicly highlighted by OpenAI.

Career Opportunities: This is primarily a disguised recruitment process. Mark Chen, OpenAI’s Chief Research Officer, confirmed that high-performing participants will be invited for interviews. The desired profile? Creative minds capable of solving novel problems, not necessarily traditional ML experts. Several current members of OpenAI’s research team are former mathematicians, physicists, or neuroscientists without formal ML backgrounds.

Real Impact: Innovations from this challenge could power OpenAI’s future embedded models. Some analysts suggest that this competition serves to crowdsource architectural breakthroughs for “Nano” and “Pico” models intended for low-power devices.