SignalTech/AI
Show HN: Maple – Sub-2-BPW Moe Pruning to Run Gemma 4 26B at 45 TPS on a 6GB GPU
Sub-2-BPW MoE pruning enables Gemma 4 26B at 45 TPS on a 6GB GPU.
Context
What happened: Aman Sachan published MAPLE, a pruning pipeline for Mixture-of-Experts LLMs, and released a compressed Gemma 4 26B-A4B checkpoint on Hugging Face.
Key numbers: Model shrinks from 9.92 GB to 4.75 GB (1.48 effective bits/param). 80 of 128 experts per layer excised, leaving 48.
Why it matters: The author states the pipeline is domain-agnostic and can be applied to other specialized fields using the repository scripts, with the released checkpoint profiled for Python programming and STEM reasoning.
What changed
Aman Sachan published MAPLE, a pruning pipeline for Mixture-of-Experts LLMs, and released a compressed Gemma 4 26B-A4B checkpoint on Hugging Face.
Why it matters
The author states the pipeline is domain-agnostic and can be applied to other specialized fields using the repository scripts, with the released checkpoint profiled for Python programming and STEM reasoning.