I Want To Enter The OpenAI x Runpod Parameter Golf Challenge With TeichAI

OpenAI and Runpod are running a challenge called Parameter Golf. The goal is simple. Build the best language model that fits in 16 megabytes and trains in under 10 minutes on 8 H100 GPUs. I want to enter this challenge with TeichAI. I have scripts that already do exactly this. I am not sure if they want to use them. That is the question.

When you have perfect scripts sitting around and a challenge that matches your skills exactly, you get excited. Then you remember you need a team. Then you ask if the team wants your scripts. This is my life.

The Challenge Details

Parameter Golf Challenge Specs

Artifact Limit: 16 MB for weights plus training code combined
Time Limit: 10 minutes on 8×H100s
Evaluation: FineWeb validation set measured in bits per byte
Timeline: March 18 through April 30, 2026
Submissions: GitHub PRs for individuals only, no teams
Compute Credits: Up to $1M available through Runpod

The challenge runs through April 30th. That gives me about two weeks. Two weeks to build the most efficient language model possible under extreme constraints. Two weeks to prove that small models can compete. Two weeks to convince TeichAI that my scripts are worth using.

What I Have

I have scripts. They work. They do exactly what the challenge requires. They fit models into 16MB. They train fast. They optimize for parameter efficiency. They are ready to go.

                        # My scripts in a nutshell

                        Step 1: Load tiny model architecture

                        Step 2: Apply aggressive quantization

                        Step 3: Train on FineWeb dataset

                        Step 4: Compress to under 16MB

                        Step 5: Submit and win (theoretically)

                        # Simple. Elegant. Probably naive.

The scripts handle everything. Tokenization. Training. Quantization. Compression. Evaluation. They are battle-tested on my Haiku models. They work on one million parameters. They should work on whatever architecture wins this challenge.

The TeichAI Question

Here is the thing. I want to enter with TeichAI. We are partners. We are building cAI-Grid together. We share the vision of efficient small models. But I do not know if they want my scripts.

Maybe they have their own approach. Maybe they want to build from scratch. Maybe they think my scripts are too complicated. Maybe they think my scripts are too simple. Maybe they have not thought about it at all.

Partnership is weird. You assume alignment. Then you realize you need to ask. Then you ask. Then you wait for an answer. This is the cycle.

Why This Matters

The challenge is about efficiency. It is about doing more with less. It is about proving that parameter constraints drive innovation. This aligns perfectly with CompactAI. This aligns perfectly with TeichAI. This aligns perfectly with everything I believe about tiny models.

The current leaderboard shows scores around 1.08 bits per byte. That is impressive. That is achievable. That is what my scripts aim for. That is what we could achieve together.

OpenAI is offering up to one million dollars in compute credits through Runpod. Quick-start level gets about twenty-five dollars. Development level gets about five hundred. Advanced level gets about one thousand. Even the advanced level covers serious training time. This removes the cost barrier. This removes the excuse.

What I Need To Do

I need to ask TeichAI. I need to show them the scripts. I need to explain why this matters. I need to convince them that together we can compete. Or I need to accept that they have their own approach. Or I need to enter alone. All outcomes are possible.

                        # My action plan

                        Step 1: Message TeichAI about the challenge

                        Step 2: Share my scripts and approach

                        Step 3: Wait for their response

                        Step 4a: If yes, collaborate and optimize

                        Step 4b: If no, enter individually or find another approach

                        Step 5: Submit before April 30

                        # Simple plan. Uncertain outcome.

The Individual Submission Rule

Here is a catch. Submissions must be individual. No teams allowed. This complicates things. If TeichAI and I work together, only one person can submit. Only one person gets credit. Only one person appears on the leaderboard.

This is fine. This is manageable. We can decide who submits. We can share the work. We can both benefit from the learning. The rule does not prevent collaboration. It just limits the credit. This is acceptable.

Final Thoughts

Parameter Golf is happening. The deadline is April 30. I have scripts. I have motivation. I have a partnership with TeichAI. I need to ask if they want to join. I need to ask if they want my scripts. I need to ask and wait and hope.

Either way, I will enter. Either with TeichAI or alone. Either with my scripts or with new ones. Either way, I will compete. Either way, I will learn. Either way, progress happens.

If you want to enter too, the challenge is open. The repo is public. The compute credits are available. The deadline approaches. The leaderboard waits for your submission. Go build something small. Go build something efficient. Go build something that fits in 16MB.