✦ AgentBrief

GPUs, VRAM tiers, mini PCs and Macs — what actually runs local models, plus daily hardware chatter.

Rough buying tiers · general guidance, Q4 quantized

8-12 GB VRAM

7-8B models (Q4)

RTX 4060 Ti 16GB, RTX 3060 12GB, Mac 16GB

Daily-driver agents, code assist, fast iteration
16-24 GB VRAM

14-32B models (Q4)

RTX 3090/4090 24GB, Mac 32-64GB

Serious local chat and agent brains
48 GB+ (2x24GB) or unified

70B models (Q4)

2x RTX 3090, Mac Studio 128-192GB

Near-frontier quality, slower tokens
None

API-first (recommended start)

Any laptop plus an API key

Zero setup, frontier models, pay per use — start here before buying hardware

Hardware chatter today

From the blog szehoyeu.github.io →