15.5× cheaper than cloud
$0.65 / 1M tokens local
Top models you can run
- SQwen 3.5 9B· 49 tok/s · 58K ctx · 9B params
- SQwen 3 8B· 56 tok/s · 63K ctx · 8B params
- SQwen 3.5 4B· 56 tok/s · 81K ctx · 4B params
Interactive tool
What should I buy for local AI?
Tell us your budget and what you'll use it for. We'll rank GPUs and Macs by how well they run models for that workload, show you which specific models will run great, and compare cost-per-token against cloud APIs.
Top 5 picks for $500 / rag / document qa
9.7× cheaper than cloud
$1.03 / 1M tokens local
Top models you can run
- SQwen 3.5 9B· 38 tok/s · 58K ctx · 9B params
- SQwen 3 8B· 43 tok/s · 63K ctx · 8B params
- SQwen 3.5 4B· 64 tok/s · 81K ctx · 4B params
23.4× cheaper than cloud
$0.43 / 1M tokens local
Top models you can run
- SQwen 3.5 4B· 56 tok/s · 54K ctx · 4B params
- SQwen 3 8B· 48 tok/s · 37K ctx · 8B params
- SPhi-4 Mini Reasoning 4B· 53 tok/s · 83K ctx · 3.799999952316284B params
19.8× cheaper than cloud
$0.50 / 1M tokens local
Top models you can run
- SQwen 3.5 4B· 56 tok/s · 52K ctx · 4B params
- SPhi-4 Mini Reasoning 4B· 53 tok/s · 80K ctx · 3.799999952316284B params
- SNemotron Nano 8B· 52 tok/s · 39K ctx · 8B params
12.3× cheaper than cloud
$0.81 / 1M tokens local
Top models you can run
- SQwen 3.5 4B· 56 tok/s · 39K ctx · 4B params
- SPhi-4 Mini Reasoning 4B· 53 tok/s · 60K ctx · 3.799999952316284B params
- SQwen 2.5 VL 7B· 98 tok/s · 33K ctx · 7B params
How we rank
We filter hardware by MSRP under your budget, then score each candidate by the best-fit model for your workload — combining fit status (native / tight / offload), decode throughput, VRAM utilization, and model quality tier. ROI uses $10/1M tokens as the cloud reference (GPT-4o / Claude Sonnet tier) with 36-month amortization and $0.15/kWh electricity. Used-market prices are flagged when applicable (e.g., RTX 3090 at ~$800).
See also