← Back to feed

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

AI · Technology · 2026-11-08 03:48

The Maple project employs sub-2-bit weight pruning to enable the Gemma 4 26B large language model to generate text at 45 tokens per second on a 6GB GPU, significantly lowering hardware requirements.

Source: github.com

Auto-generated summary for reference only. RSS