ModelDock
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 75% reading / 25% 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.
| Hardware | Capability | Economics for your workload | Evidence | Market | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 309024GB · single-card node | 24 GB | 705 | 55.9 | Good | 500W | $1,350 | $2,350 | 15.6 | $0.198 | $1.80 | $2.73 | 2,524 d | Median ↗ | |
| RTX 3090 Ti24GB · single-card node | 24 GB | 1,458.3 | 69.8 | Good | 600W | $1,400 | $2,400 | 21.1 | $0.165 | $2.16 | $3.69 | 1,570 d | Median ↗ | |
| AMD Radeon RX 7900 XTX24GB GDDR6 | 24 GB | 877.8 | 43.9 | Good | 475W | $800 | $1,800 | 13.2 | $0.204 | $1.71 | $2.31 | 3,007 d | Median ↗ | |
| 2× RTX 5060 Ti16GB each · dual-card node | 32 GB | 723 | 65.9 | Good | 480W | $1,760 | $2,760 | 17.9 | $0.181 | $1.73 | $3.13 | 1,969 d | Median ↗ | |
| 2× RTX 5070 Ti16GB each · dual-card node | 32 GB | 3,250 | 113.3 | Good | 620W | $2,400 | $3,400 | 35.4 | $0.116 | $2.23 | $6.20 | 857 d | Median ↗ | |
| 2× RTX 408016GB each · dual-card node | 32 GB | ~1,700 | ~58 | Good | 760W | $1,900 | $2,900 | 18.2 | $0.238 | $2.74 | $3.18 | 6,501 d | Estimate | |
| RTX 409024GB · single-card node | 24 GB | 2,692.3 | 68.8 | Good | 550W | $2,500 | $3,500 | 22.1 | $0.176 | $1.98 | $3.86 | 1,858 d | Median ↗ | |
| RTX 509032GB · single-card node | 32 GB | 3,571.4 | 132.5 | Good | 700W | $4,200 | $5,200 | 41.2 | $0.130 | $2.52 | $7.21 | 1,109 d | Median ↗ | |
| Intel Arc Pro B7032GB GDDR6 | 32 GB | 424.8 | 49.3 | Good | 350W | $1,299 | $2,299 | 12.6 | $0.199 | $1.26 | $2.21 | 2,412 d | Median ↗ | |
| RTX PRO 5000 Blackwell48GB GDDR7 ECC | 48 GB | 6,288 | 112.5 | Good | 420W | $9,085 | $10,085 | 36.9 | $0.191 | $1.51 | $6.46 | 2,040 d | Median ↗ | |
| RTX PRO 6000 Blackwell96GB GDDR7 ECC | 96 GB | 7,756 | 109.3 | Good | 720W | $13,250 | $14,250 | 36.2 | $0.287 | $2.59 | $6.34 | 3,803 d | Median ↗ | |
| Mac Studio M3 Ultra 512 GB80-core · 512GB unified memory | 512 GB | 462.5 | 44.3 | Good | 330W | — | $9,499 | 11.9 | $0.538 | $1.19 | $2.08 | 10,657 d | Median ↗ | |
| Mac Studio M4 Max 64 GB64GB unified memory · current retail reference | 64 GB | ~274 | ~53.3 | Good | 145W | — | $3,499 | 11.6 | $0.210 | $0.52 | $2.04 | 2,312 d | Estimate | |
| Mac Studio M4 Max 128 GBsource-backed oMLX profile | 128 GB | — | 56.8 | Good | 145W | — | $6,500 | — | pending | — | — | — | Median ↗ | |
| MacBook Pro M5 Max 128 GB128GB unified memory | 128 GB | 745 | 59.2 | Good | 120W | — | $6,699 | 16.5 | $0.248 | $0.43 | $2.89 | 2,725 d | Median ↗ | |
| Mac mini M5 Pro 64 GB20-core GPU · 64GB unified memory | 64 GB | ~373 | ~29.6 | Good | ~110W | — | $2,899 | 8.3 | $0.240 | $0.40 | $1.45 | 2,762 d | Estimate | |
| Mac Studio M5 Max 64 GB40-core GPU · 64GB unified memory | 64 GB | ~745 | ~59.2 | Good | ~180W | — | $3,799 | 16.5 | $0.165 | $0.65 | $2.89 | 1,694 d | Estimate | |
| Mac Studio M5 Max 128 GB40-core GPU · 128GB unified memory | 128 GB | ~745 | ~59.2 | Good | ~180W | — | $5,399 | 16.5 | $0.218 | $0.65 | $2.89 | 2,407 d | Estimate | |
| Mac Studio M5 Ultra 96 GB30-core CPU · 64-core GPU · 96GB unified memory | 96 GB | ~1,075 | ~91 | Good | ~360W | — | $5,499 | 25.1 | $0.172 | $1.30 | $4.39 | 1,778 d | Estimate | |
| Mac Studio M5 Ultra 256 GB36-core CPU · 80-core GPU · 256GB unified memory | 256 GB | ~1,075 | ~91 | Good | ~360W | — | $10,799 | 25.1 | $0.288 | $1.30 | $4.39 | 3,492 d | Estimate | |
| DGX Spark128GB unified memory | 128 GB | 1,225.5 | 44 | Good | 170W | — | $4,699 | 13.7 | $0.232 | $0.61 | $2.40 | 2,623 d | Median ↗ | |
capability efficiency cost ~ hardware-scaled rate estimate or incomplete evidence
Assumptions
75% reading / 25% writing · 24 active h/day
Whole-node power · Qwen3.8 27B · ordinary local use
Transparent calculation
total GPU rig = GPU price + $1,000 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 $1,000 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.