How much VRAM do you really need to run 13B models locally?
close_Meal6005 · reddit · 2026-08-25
The poster is looking to buy a dedicated GPU for local models and found conflicting community advice: some say 8GB VRAM is enough to run 13B models, others insist on 16GB minimum. He asks for real-world experience—whether 8GB actually works or becomes annoyingly slow.
More from Infra
- IBM unveils 2nm dual-architecture processor running ARM and IBM Z instructions concurrently — bookwormengr · 2026-08-25
- Nvidia Vera CPU exec calls agentic AI the most complex computing workload in history — firstadopter · 2026-08-25
- Hyperscale Data Centers Use 1.5GWh/Day vs 50GWh for Steel — davidpattersonx · 2026-08-25
- NVIDIA blog: Gemma 4 hits 10,996 OTSU with Vera Rubin optimizations — ricklamers · 2026-08-25
- sPTC speeds up agents via speculative tool calling — a1zhang · 2026-08-25
- Speculative Programmatic Tool Calling Overlaps Code Gen and LLM Inference — a1zhang · 2026-08-25