Hardware benchmarks
Ordinary-lane medians for Qwen3.8 Flash Next. Local cost and break-even follow the hours, electricity, and API prices you set.
Qwen3.8 Flash Next · one stream · read, then write · ModelDock benchmarks
Default API prices are $0.10 per million reading tokens and $0.64 per million writing tokens. At a 90/10 mix that blend matches the published API-value column. Change either price to match the API you actually pay.
Your workload
29 hardware profiles · Dual-GPU rows need chassis clearance, power headroom, and multi-GPU scheduling.
| Fit | Evidence | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 4090 single-card node | 24 GB | 4,000 | 109.3 | Good | 550 W | $4,199 | 75.8 | $0.056 | $1.98 | $11.67 | 433 d | Median |
| RTX 5080 16GB GDDR7 · single-card node | 16 GB | 3,712.5 | 80.7 | Good | 510 W | $2,800 | 58.3 | $0.058 | $1.84 | $8.98 | 392 d | Median |
| RTX 5070 Ti 16GB · Q3 or smaller only | 16 GB | 2,498.5 | 71.1 | Good | 450 W | $2,380 | 48.9 | $0.060 | $1.62 | $7.53 | 403 d | Median |
| RTX 3090 single-card node | 24 GB | 2,094 | 86.5 | Good | 500 W | $2,700 | 54.5 | $0.060 | $1.80 | $8.39 | 410 d | Median |
| RTX 4070 Ti SUPER 16GB 16GB GDDR6X · single-card node | 16 GB | 1,700 | 72 | Good | 435 W | $2,140 | 45.0 | $0.061 | $1.57 | $6.94 | 398 d | Median |
| RTX 5090 single-card node | 32 GB | 5,803 | 168.4 | Good | 700 W | $8,500 | 115.4 | $0.062 | $2.52 | $17.77 | 558 d | Median |
| RTX 3090 Ti single-card node | 24 GB | ~2,255.1 | ~93.1 | Good | ~600 W | $2,774 | 58.6 | $0.063 | $2.16 | $9.03 | 404 d | Estimate |
| AMD Radeon RX 7900 XTX 24GB GDDR6 · single-card node | 24 GB | 1,459.3 | 75.5 | Good | 475 W | $2,000 | 44.5 | $0.063 | $1.71 | $6.85 | 389 d | Median |
| AMD Radeon RX 9070 XT 16GB 16GB GDDR6 · single-card node | 16 GB | 1,652 | 53.8 | Good | 454 W | $2,000 | 35.9 | $0.076 | $1.63 | $5.54 | 513 d | Median |
| Ryzen AI Max+ 395 128 GB Radeon 8060S · 128GB unified memory | 128 GB | 1,320 | 51.4 | Good | 170 W | $3,500 | 32.9 | $0.077 | $0.61 | $5.06 | 786 d | Median |
| 2× RTX 5060 Ti 16GB each · dual-card node | 32 GB | 1,602 | 67.6 | Good | 480 W | $2,960 | 42.3 | $0.079 | $1.73 | $6.52 | 618 d | Median |
| RTX 5070 12 GB 12GB GDDR7 · single-card node | 12 GB | 1,443.2 | 48.3 | Good | 400 W | $2,070 | 32.1 | $0.080 | $1.44 | $4.94 | 592 d | Median |
| 2x RTX 3090 24 GB 24 GB each · dual-card layer split | 48 GB | 2,137.6 | 111.3 | Tight | 850 W | $4,200 | 65.5 | $0.082 | $3.06 | $10.08 | 598 d | Median |
| Mac Studio M5 Ultra 96 GB 30-core CPU · 64-core GPU · 96GB unified memory | 96 GB | ~2,284.8 | ~64.8 | Good | ~360 W | $5,499 | 44.6 | $0.097 | $1.30 | $6.87 | 987 d | Estimate |
| DGX Spark 128GB unified memory | 128 GB | 2,228.5 | 48.7 | Good | 170 W | $5,100 | 35.2 | $0.097 | $0.61 | $5.41 | 1,062 d | Median |
| RTX PRO 6000 Blackwell 96GB GDDR7 ECC · single-card node | 96 GB | 8,987.9 | 171.1 | Good | 720 W | $18,317 | 126.2 | $0.100 | $2.59 | $19.44 | 1,087 d | Median |
| 2x RTX 3060 12 GB 12 GB each · dual-card layer split | 24 GB | 1,239.8 | 41.3 | Good | 490 W | $1,814 | 27.5 | $0.100 | $1.76 | $4.23 | 736 d | Median |
| MacBook Pro M5 Max 128 GB 128GB unified memory | 128 GB | 1,310 | 67.3 | Good | 120 W | $6,699 | 39.8 | $0.103 | $0.43 | $6.12 | 1,177 d | Median |
| 2x TITAN RTX 24 GB 24 GB each · dual-card layer split | 48 GB | 1,698.2 | 60.6 | Good | 710 W | $2,881 | 39.6 | $0.104 | $2.56 | $6.10 | 812 d | Median |
| AMD Radeon RX 6900 XT 16GB 16GB GDDR6 · single-card node | 16 GB | 466 | 47.6 | Good | 450 W | $1,669 | 21.4 | $0.118 | $1.62 | $3.30 | 994 d | Median |
| 2x AMD Radeon RX 6900 XT 16GB 16 GB each · dual-card layer split | 32 GB | 772.3 | 63 | Good | 750 W | $2,138 | 31.4 | $0.123 | $2.70 | $4.83 | 1,002 d | Median |
| Mac Studio M5 Ultra 256 GB 36-core CPU · 80-core GPU · 256GB unified memory | 256 GB | 2,856 | 81 | Good | 360 W | $10,799 | 55.8 | $0.129 | $1.30 | $8.59 | 1,481 d | Median |
| Mac Studio M5 Max 64 GB 40-core GPU · 64GB unified memory | 64 GB | ~517.9 | ~36 | Good | ~180 W | $3,799 | 19.1 | $0.143 | $0.65 | $2.95 | 1,653 d | Estimate |
| MacBook Pro M5 Max 64 GB 40-core GPU · 64GB unified memory · source chassis undisclosed | 64 GB | 517.9 | 36 | Good | 120 W | $5,099 | 19.1 | $0.169 | $0.43 | $2.95 | 2,028 d | Median |
| Mac Studio M5 Max 128 GB 40-core GPU · 128GB unified memory | 128 GB | ~517.9 | ~36 | Good | ~180 W | $5,399 | 19.1 | $0.188 | $0.65 | $2.95 | 2,349 d | Estimate |
| Mac mini M5 Pro 64 GB 18-core CPU · 20-core GPU · 64GB unified memory | 64 GB | ~259 | ~18 | Good | ~110 W | $2,899 | 9.6 | $0.207 | $0.40 | $1.47 | 2,691 d | Estimate |
| Mac Studio M4 Max 128 GB 128GB unified memory · source-backed oMLX profile | 128 GB | 446.8 | 32.5 | Good | 145 W | $6,500 | 17.0 | $0.241 | $0.52 | $2.61 | 3,108 d | Median |
| Mac Studio M3 Ultra 512 GB 80-core · 512GB unified memory | 512 GB | 479.6 | 18.1 | Good | 330 W | $9,499 | 11.7 | $0.548 | $1.19 | $1.80 | 15,580 d | Median |
| RTX 4070 12 GB 12GB board-dependent GDDR6/GDDR6X · single-card node | 12 GB | 1,376 | 36.9 | Good | 350 W | — | 25.7 | — | $1.26 | $3.96 | — | Median |
Selected rig
RTX 4090
24 GB · single-card node
Reading tok/s
4,000Writing tok/s
109.3Power
550 WM tok/day
75.8Local $/M
$0.056Energy/day
$1.98API value/day
$11.67Break-even
433 dHow the numbers move
- Daily capacity is active hours times the blended rate. Time is sequential: reading share divided by read tok/s, plus writing share divided by write tok/s.
- Energy per day is whole-node watts times hours times your electricity price.
- Local cost per million tokens is energy plus hardware spread over the life you set, divided by daily capacity. A discrete GPU rig in this table is the card price plus a $1,200 host.
- Break-even days are the rig price divided by API value per day minus energy per day. That is a comparison with the API prices above, not a promise of savings.