Researchers compress a 24-million parameter language model into just 15MB using GPTQ-lite and Muon optimization techniques

GolfStudent v2 achieves extreme model compression, reducing a 24M-parameter LLM to only 15MB in size

Uses GPTQ-lite quantization combined with Muon optimization to achieve the compression

Contribution submitted to OpenAI's parameter-golf project on GitHub, focusing on efficient model design

Demonstrates significant progress in making capable language models deployable on resource-constrained devices