Default path
Qwen3.8 27B under local, single-user, single-stream use with stable MTP/speculative draft treated as the normal path. Source-marked short-context probes, SGLang, and DFlash/DFlash2 stay outside the default median.
ModelDock research
Benchmark Data
Ordinary-lane hardware · source medians · adjustable economicsYour workload
Reading is the model taking in your prompt, files, and chat history. Writing is the answer it generates. Long documents and coding sessions are usually read-heavy, so the default starts at 90% reading / 10% writing. Stable MTP is treated as normal, not a separate scenario.
Model Quantization
2× GPU build note Dual-card rigs require chassis clearance and slot spacing for wide cards, stable high-headroom power delivery and airflow, plus more involved installation and multi-GPU scheduling than a single-card node.
| Capability | Economics for your workload | Evidence | Market | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Mac Studio M5 Ultra 96 GB30-core CPU · 64-core GPU · 96GB unified memory | 96 GB | 4,083 | 139.6 | Good | ~360W | — | $5,499 | 92.2 | $0.047 | $1.30 | $14.20 | 427 d | Median ↗ | |
| RTX 409024GB · single-card node | 24 GB | 4,000 | 109.3 | Good | 550W | $3,199 | $4,399 | 75.8 | $0.058 | $1.98 | $11.67 | 454 d | Median ↗ | |
| RTX 508016GB GDDR7 · single-card node | 16 GB | 3,591.4 | 80.1 | Good | ~510W | $1,600 | $2,800 | 57.6 | $0.059 | $1.84 | $8.87 | 398 d | Median ↗ | |
| AMD Radeon AI PRO R970032GB GDDR6 · single-card node | 32 GB | 2,197.4 | 80.8 | Good | ~420W | $1,700 | $2,900 | 52.5 | $0.059 | $1.51 | $8.08 | 442 d | Median ↗ | |
| RTX 309024GB · single-card node | 24 GB | 2,094 | 86.5 | Good | 500W | $1,399 | $2,599 | 54.5 | $0.059 | $1.80 | $8.39 | 395 d | Median ↗ | |
| RTX 509032GB · single-card node | 32 GB | 6,256.6 | 173.8 | Good | 700W | $7,250 | $8,450 | 120.1 | $0.060 | $2.52 | $18.50 | 529 d | Median ↗ | |
| RTX 5070 Ti16GB · Q3 or smaller only | 16 GB | 2,498.5 | 71.1 | Good | 450W | $1,235.5 | $2,435.5 | 48.9 | $0.060 | $1.62 | $7.53 | 413 d | Median ↗ | |
| RTX 4070 Ti SUPER 16GB16GB GDDR6X · single-card node | 16 GB | 1,700 | 72 | Good | ~435W | $920 | $2,120 | 45 | $0.061 | $1.57 | $6.94 | 395 d | Median ↗ | |
| 2× RTX 5070 Ti16GB each · dual-card node | 32 GB | 3,030 | 96 | Good | 620W | $2,471 | $3,671 | 64.5 | $0.066 | $2.23 | $9.94 | 477 d | Median ↗ | |
| RTX 3090 Ti24GB · single-card node | 24 GB | ~2,255.1 | ~93.1 | Good | 600W | $1,900 | $3,100 | 58.6 | $0.066 | $2.16 | $9.03 | 452 d | Estimate | |
| AMD Radeon RX 7900 XTX24GB GDDR6 | 24 GB | 1,529.2 | 71.2 | Good | 475W | $1,200 | $2,400 | 43.4 | $0.070 | $1.71 | $6.68 | 484 d | Median ↗ | |
| 2× Tesla V100 32GB32GB each · dual-card PCIe node | 64 GB | 2,355.3 | 78.8 | Good | ~650W | $1,500 | $2,700 | 52.3 | $0.073 | $2.34 | $8.06 | 473 d | Median ↗ | |
| AMD Radeon RX 9070 XT 16GB16GB GDDR6 · single-card node | 16 GB | 1,660 | 54.9 | Good | ~454W | $800 | $2,000 | 36.6 | $0.075 | $1.63 | $5.63 | 501 d | Median ↗ | |
| Mac Studio M5 Ultra 256 GB36-core CPU · 80-core GPU · 256GB unified memory | 256 GB | 3,441.5 | 155 | Good | ~360W | — | $10,799 | 95.3 | $0.076 | $1.30 | $14.68 | 808 d | Median ↗ | |
| 2× RTX 5060 Ti16GB each · dual-card node | 32 GB | 1,602 | 67.6 | Good | 480W | $1,760 | $2,960 | 42.3 | $0.079 | $1.73 | $6.52 | 618 d | Median ↗ | |
| RTX 5070 12 GB12GB GDDR7 · single-card node | 12 GB | 1,443.2 | 48.3 | Good | ~400W | $835 | $2,035 | 32.1 | $0.080 | $1.44 | $4.94 | 582 d | Median ↗ | |
| 2x RTX 3090 24 GB24 GB each · dual-card layer split | 48 GB | 2,137.6 | 111.3 | Tight | ~850W | $2,798 | $3,998 | 65.5 | $0.080 | $3.06 | $10.08 | 570 d | Median ↗ | |
| Mac Studio M5 Max 128 GB40-core GPU · 128GB unified memory | 128 GB | 2,146.5 | 65.2 | Good | ~180W | — | $5,399 | 44.3 | $0.081 | $0.65 | $6.82 | 876 d | Median ↗ | |
| Ryzen AI Max+ 395 128 GBRadeon 8060S · 128GB unified memory | 128 GB | 1,242.7 | 47.5 | Good | ~170W | — | $3,500 | 30.5 | $0.083 | $0.61 | $4.70 | 856 d | Median ↗ | |
| 2× AMD Radeon AI PRO R9700 32 GB32 GB each · dual-card node · card-set price derived: 2 × $1,699.99; no observed pair listing | 64 GB | 4,244.5 | 79.2 | Good | ~750W | $3,400 | $4,600 | 58.6 | $0.089 | $2.70 | $9.02 | 729 d | Median ↗ | |
| Intel Arc Pro B7032GB GDDR6 | 32 GB | 782.8 | 51.4 | Good | 350W | $1,300 | $2,500 | 27.9 | $0.094 | $1.26 | $4.30 | 823 d | Median ↗ | |
| MacBook Pro M5 Max 128 GB128GB unified memory | 128 GB | 1,245.5 | 77.6 | Good | 120W | — | $6,699 | 43 | $0.096 | $0.43 | $6.62 | 1,084 d | Median ↗ | |
| DGX Spark128GB unified memory | 128 GB | 2,228.5 | 48.7 | Good | 170W | — | $5,100 | 35.2 | $0.097 | $0.61 | $5.41 | 1,062 d | Median ↗ | |
| 2x TITAN RTX 24 GB24 GB each · dual-card layer split | 48 GB | 1,713.4 | 66.1 | Good | ~710W | $1,681 | $2,881 | 42.4 | $0.098 | $2.56 | $6.53 | 726 d | Median ↗ | |
| Intel Arc Pro B60 24GB24GB GDDR6 · single-card node | 24 GB | 1,034 | 35.5 | Good | ~350W | $750 | $1,950 | 23.4 | $0.099 | $1.26 | $3.61 | 831 d | Median ↗ | |
| 2x RTX 3060 12 GB12 GB each · dual-card layer split | 24 GB | 1,239.8 | 41.3 | Good | ~490W | $614 | $1,814 | 27.4 | $0.101 | $1.76 | $4.22 | 738 d | Median ↗ | |
| RTX PRO 6000 Blackwell96GB GDDR7 ECC | 96 GB | 8,987.9 | 171.1 | Good | 720W | $18,000 | $19,200 | 126.2 | $0.104 | $2.59 | $19.43 | 1,140 d | Median ↗ | |
| AMD Radeon RX 6900 XT 16GB16GB GDDR6 · single-card node | 16 GB | 466 | 47.6 | Good | ~450W | $469 | $1,669 | 21.4 | $0.118 | $1.62 | $3.30 | 994 d | Median ↗ | |
| 2x AMD Radeon RX 6900 XT 16GB16 GB each · dual-card layer split | 32 GB | 772.3 | 63 | Good | ~750W | $938 | $2,138 | 31.4 | $0.123 | $2.70 | $4.83 | 1,002 d | Median ↗ | |
| RTX 3060 12 GB12 GB GDDR6 · single-card node; host RAM offload | 12 GB | 820 | 20.2 | Good | ~320W | $360 | $1,560 | 14.3 | $0.140 | $1.15 | $2.20 | 1,489 d | Median ↗ | |
| Mac Studio M5 Max 64 GB40-core GPU · 64GB unified memory | 64 GB | ~517.9 | ~36 | Good | ~180W | — | $3,799 | 19.1 | $0.143 | $0.65 | $2.95 | 1,653 d | Estimate | |
| Mac mini M5 Pro 64 GB18-core CPU · 20-core GPU · 64GB unified memory | 64 GB | 430 | 19.2 | Good | ~110W | — | $2,899 | 11.8 | $0.168 | $0.40 | $1.82 | 2,030 d | Median ↗ | |
| MacBook Pro M5 Max 64 GB40-core GPU · 64GB unified memory · source chassis undisclosed | 64 GB | 517.9 | 36 | Good | ~120W | — | $5,099 | 19.1 | $0.169 | $0.43 | $2.95 | 2,028 d | Median ↗ | |
| Mac Studio M4 Max 128 GBsource-backed oMLX profile | 128 GB | 446.8 | 32.5 | Good | 145W | — | $6,500 | 17 | $0.241 | $0.52 | $2.61 | 3,111 d | Median ↗ | |
| Mac Studio M3 Ultra 512 GB80-core · 512GB unified memory | 512 GB | 479.6 | 18.1 | Good | 330W | — | $9,499 | 11.7 | $0.548 | $1.19 | $1.80 | 15,577 d | Median ↗ | |
| RTX 4070 12 GB12GB board-dependent GDDR6/GDDR6X · single-card node | 12 GB | 1,376 | 36.9 | Good | ~350W | pending | — | 25.7 | pending | — | — | — | Median ↗ | |
| RTX 4080 16GB16GB GDDR6X · stock single-card node; host RAM separate | 16 GB | 1,996.3 | 66 | Good | ~470W | pending | — | 43.9 | pending | — | — | — | Median ↗ | |
capability efficiency cost ~ hardware-scaled rate estimate or incomplete evidence
Assumptions
90% reading / 10% writing · 24 active h/day
Whole-node power · Qwen3.8 27B · ordinary local use
Transparent calculation
total GPU rig = GPU price + $1,200 hostlocal $/MM = (energy + hardware allocation) / daily capacitybreak-even days = total rig / (API value/day − energy/day)Sequential single-user model: reading time + writing time; no concurrency or overlap assumed.
Evidence data
Open the source spread behind published hardware records.
Methodology
The conditions behind the table.
Performance uses a source-visible MTP-ready median. The economics below are adjustable because they describe your usage—not the benchmark run.
Per-metric median
Reading (prefill) and writing (decode) are medianed independently. Source conditions—quant, context, runtime, topology, and draft depth—remain visible after clicking Median.
Evidence labels
Median links to the source spread. ~ Rate estimate uses a disclosed conservative hardware ratio derived from measured and reported reading and writing data; it remains separate from source medians. Pending means no ordinary-lane source metric has been collected yet.
Pricing basis
Discrete GPU totals use the dated GPU price plus a $1,200 host/platform allowance. Break-even divides total rig price by equivalent API value per day less direct energy cost; it is not an investment-return guarantee.