Dev Prunes 85% of Gemma 4 26B Weights, Keeps 92% Agreement at 2x Speed
Accomplished_Ad_56 · reddit · 2026-11-04
A developer pruned Gemma 4 26B-A4B (MoE) down to a dense 3.81B model by keeping 8 experts per layer. On a narrow classification task it achieves 91.7% agreement with the original 26B (vs 90% for Gemma 4 12B), runs at 79 tokens/s on Mac (2x the 12B) using only 3.8 GB memory — at the cost of losing broad knowledge.
Training used block-local training on self-collected data, optimizing for "same thought process" rather than same answers. The method is efficient enough to retrain on-device with 5 GB memory to adapt to deployment data. Full details in the linked blog post.
More from Infra
- citrini: being part of AI supply chain no longer guarantees 100% yearly stock gains — citrini · 2026-11-05
- Building orbital data centers to dodge regulators is 'completely unserious' — zacharynado · 2026-11-05
- Verda Launches Inference Preview with Day-0 Access to Kimi K3.1 Pro — Kimi_Moonshot · 2026-11-05
- Banach: open-source engine streams 1T-parameter Kimi K2 from NVMe on a 64GB desktop — gautamrbharadwaj · 2026-11-05
- Google unveils $10B, 200-hectare datacenter project in Châteauroux, France with up to 1.2GW — QuintinPope5 · 2026-11-05
- Corebrum OS adds built-in VNC screen share, demoed remote-viewing an NVIDIA Thor robot from a Mac — chrismatthieu · 2026-11-05