A 35B MoE on a 16 GB GPU, without the offload tax

Keyword: Cache

Publisher: Lucebox.com

Published: Jun 8, 2026

June 2026 By Davide Ciffa A 33-35B mixture-of-experts model fires only a handful of its experts per token, but to keep it on the GPU you still pay for all of them. Luce Spark pins the experts your … [+9465 chars]

Read Full Story ↗