Show HN: Maple – Sub-2-BPW Moe Pruning to Run Gemma 4 26B at 45 TPS on a 6GB GPU

Aman Sachan, an independent researcher, has developed a pruning technique called MAPLE that reduces the size of a 26B parameter large language model, Gemma 4 26B-A4B, from 9.92 GB to 4.75 GB, allowing it to run entirely within a 6 GB GPU on a laptop. This is achieved by excising 80 dormant experts from the model, while preserving its ability to perform deductive reasoning and syntax generation. The technique is domain-agnostic and can be applied to any specialized field, and the compressed model is publicly hosted on Hugging Face.