Rig Finder index / 32B / Local RAG / knowledge base / Washington
Best rig to run a 32B model for local rag / knowledge base in Washington
For a 32B model at this workload, the RTX 3090 (used) pays off fastest — break-even ~6 months, keeping $6,219 over 3 years versus Claude Sonnet 4.6.
Same workload, same cloud, same electricity — we swap only the hardware and rank by real break-even. Rigs that can't run a 32B model are shown, never recommended. Compared against Claude Sonnet 4.6 ($3/M in, $15/M out); change the cloud in the lab.
Every build, ranked
| Rig | Build | 32B tok/s | Local/mo | Break-even | Verdict |
|---|---|---|---|---|---|
| RTX 3090 (used)best payoff | $1,200 | 24estimated | $43.25 | ~6 mo | OWN |
| RTX 4090 | $2,300 | 38estimated | $74.88 | ~11 mo | OWN |
| 2× RTX 3090 | $2,250 | 30estimated | $79.28 | ~11 mo | OWN |
| Mac M-series | $3,000 | 20estimated | $85.32 | ~14 mo | OWN |
| RTX 5090 | $3,500 | 60estimated | $111 | ~17 mo | OWN |
| Mini-PCtoo small for 32B | $750 | — | $22.98 | ~4 mo | OWN |
Assumes local rag / knowledge base use: ~40M tokens/month, 5h/day active, 11h/day idle, at Washington electricity ($0.118/kWh). Speeds are our estimates; community-measured tok/s live on the interactive Rig Finder and the benchmarks board. Every formula is on the methodology page.
Run this with your exact numbers →