Six rigs · rev. 2026.10
Builds
Six illustrative rigs with parts lists, estimated power and speed, and cost comparisons. Sample prices need checking before purchase.
Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology
Illustrative build · sample prices
Value 16 GB
Complete system, listed parts only; tax, shipping and peripherals excluded.
- Qwen3 14B
- 26.0 tok/s (estimate)
- Largest estimated fit
- gpt-oss-20b
- Assumed load power
- 250 W
- Illustrative break-even, $200 plan
- 6.1 mo
Power assumption: 250 W × 4 h load + 45 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
A new 16 GB card in a plain AM5 tower. Runs every 14B model at Q4 with long context, and gpt-oss-20b at full speed.
Illustrative build · sample prices
Used-3090 24 GB
Complete system, listed parts only; tax, shipping and peripherals excluded.
- Gemma 3 27B
- 31.8 tok/s (estimate)
- Largest estimated fit
- Qwen3 32B
- Assumed load power
- 420 W
- Illustrative break-even, $200 plan
- 10.6 mo
Power assumption: 420 W × 4 h load + 70 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
The reference rig. A used RTX 3090 gives 24 GB at 936 GB/s, enough for every 24–32B model at Q4.
Illustrative build · sample prices
Mac Studio M5 Max
Complete system, listed parts only; tax, shipping and peripherals excluded.
- Gemma 3 27B
- 11.7 tok/s (estimate)
- Largest estimated fit
- Qwen3 32B
- Assumed load power
- 150 W
- Illustrative break-even, $200 plan
- 12.7 mo
Power assumption: 150 W × 4 h load + 12 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
No build, no noise. The base Mac Studio runs 27B-class models at Q4 and sits silently on a desk.
Illustrative build · sourced prices
Strix Halo SFF 128 GB
Complete system, listed parts only; tax, shipping and peripherals excluded.
- gpt-oss-120b
- 37.5 tok/s (estimate)
- Largest estimated fit
- gpt-oss-120b
- Assumed load power
- 120 W
- Illustrative break-even, $200 plan
- 17.8 mo
Power assumption: 120 W × 4 h load + 15 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
128 GB of unified memory in a mini PC. Holds gpt-oss-120b with room to spare; dense 70B models are slow.
Illustrative build · sample prices
Dual-GPU 48 GB
Complete system, listed parts only; tax, shipping and peripherals excluded.
- Llama 3.3 70B
- 12.4 tok/s (estimate)
- Largest estimated fit
- Llama 3.3 70B
- Assumed load power
- 780 W
- Illustrative break-even, $200 plan
- 23.0 mo
Power assumption: 780 W × 4 h load + 110 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
Two used 3090s split a 70B model by layers. Capacity doubles; single-stream bandwidth does not.
Illustrative build · sample prices
Silent home server
Complete system, listed parts only; tax, shipping and peripherals excluded.
- Qwen3 8B
- 27.7 tok/s (estimate)
- Largest estimated fit
- gpt-oss-20b
- Assumed load power
- 230 W
- Illustrative break-even, $200 plan
- 6.6 mo
Power assumption: 230 W × 4 h load + 30 W × 20 h idle daily, 30 days at $0.15/kWh. Cost comparison only; capabilities differ.
A low-power 16 GB box for an always-on household assistant. Tuned for idle draw, not peak speed.