VRAM Tier Testing: 10GB VRAM + 32GB RAM Runs 35B Sparse Model

OutdatedMemeKing · reddit · 2026-08-08

A developer shares a hardware configuration table based on real testing for local AI apps: no GPU runs a small chat model (2.6GB); 8GB runs chat+image+tools (5.1GB); 10-12GB adds a vision model (11.1GB); 24GB holds large model, vision, and code simultaneously (20.6GB). Surprising finding: 10GB VRAM + 32GB system RAM can run a 35B sparse model (27GB), slow but usable. Also notes model swapping matters more than model size.

Original post →

More from Embodied

Embodied channel →