Parameter Golf fits 24M LLM in 15MB

At this scale, quantization is the product, not a finishing touch. The submission reads like GPTQ-lite plus architecture tuning, which is exactly the kind of compression-first work that moves a tiny-model leaderboard.

• –Five clip candidates per row is a very cheap search space, but it directly attacks reconstruction error instead of assuming one clipping heuristic will be good enough.
• –The width-over-depth result suggests that once the byte ceiling is tight, representational breadth matters more than stacking layers.
• –Landing around 15.06MB means the margin is microscopic, so schedule tweaks, optimizer choices, and packing format can all move the ranking.
• –Because OpenAI's rules fix the dataset and cap training time, leaderboard separation is likely coming from many small compounding improvements, not one dramatic breakthrough.