Marketing sells you FPS and ray tracing. Local AI cares about exactly one spec: VRAM capacity. This guide turns that into a shopping decision you won’t regret.
The only rule
A model at Q4 quantization needs roughly 0.6GB per billion parameters, plus 1–2GB for context and overhead. Memorize that and every product page starts making sense:
- 8B model → ~6GB → fits 8GB cards (barely), 12GB cards (comfortably)
- 14B model → ~10GB → 12GB cards (tight), 16GB cards (comfortable)
- 32B model → ~20GB → 24GB cards
- 70B model → ~43GB → dual 24GB or Mac unified memory
The three smart buys
Used RTX 3060 12GB (~$250–300). The community’s first-rig classic. Runs 8B–14B models, SDXL, Whisper. Boring, abundant, correct.
Used RTX 3090 24GB (~$700–900). The best VRAM-per-dollar in history. Runs 32B reasoning models and FLUX. Buy from reputable resellers, test under load on day one.
16GB mid-range (RTX 4060 Ti 16GB / 5060 Ti 16GB, ~$450–500). The new-value pick with warranty, modern architecture and enough room for Qwen3 14B with real context.
Mac exception
Apple silicon changes the math: unified memory means a 36GB MacBook or Mac Studio loads 32B models that no 24GB PC can touch. You pay for convenience and silence, and image generation is slower than NVIDIA. If you’re already in the ecosystem, don’t buy a PC just for LLMs.
Red flags on the used market
- Mining-farm cards with no receipt (ask for original purchase proof)
- Prices below market by 30%+ (scams and dead memory chips)
- Blower-style cards from servers: fine chips, but check fan bearings
Verify before you trust
Any listing: run matmul or a local model for 30 minutes under load. Temperatures under 85°C and no artifacts = green light.
Where to next
- Compare tiers side by side in the Local AI hub.
- Run the VRAM calculator with the models you actually want.
Some links are affiliate links. They never change our rankings or the math — they help keep AIVerseWeb independent.