Home / GPU ladder / RTX PRO 6000 Blackwell 96GB

RTX PRO 6000 Blackwell 96GB for MiniMax H3: Speed, VRAM & Recommended Setup

96GB GDDR7 · bandwidth 1792 GB/s · Blackwell · sm_120 · estimated with 32GB RAM / fast-disk · updated 2026-09-28

Verdict: No workable config with the default combo — lower the resolution or switch to a smaller quant (see the model library).

Size (16:9)	5s	8s	10s	15s
416 × 256 0.1MP · flicker-prone	✕ 1 min 24 s VRAM 69G	✕ 1 min 35 s VRAM 69G	✕ 1 min 44 s VRAM 69G	✕ 2 min 6 s VRAM 70G
864 × 480 0.4MP	✕ 2 min 26 s VRAM 70G	✕ 3 min 32 s VRAM 71G	✕ 4 min 33 s VRAM 72G	✕ 7 min 54 s VRAM 74G
960 × 544 0.5MP	✕ 2 min 53 s VRAM 71G	✕ 4 min 35 s VRAM 72G	✕ 6 min 8 s VRAM 73G	✕ 11 min 27 s VRAM 76G
1152 × 640 0.7MP	✕ 4 min VRAM 72G	✕ 7 min 19 s VRAM 74G	✕ 10 min 22 s VRAM 75G	✕ 21 min 26 s VRAM 79G
1344 × 768 1.0MP · official max	✕ 6 min 6 s VRAM 73G	✕ 12 min 37 s VRAM 76G	✕ 18 min 52 s VRAM 78G	✕ 42 min 44 s VRAM 83G
1376 × 768 1.0MP · above native	✕ 6 min 19 s VRAM 73G	✕ 13 min 9 s VRAM 76G	✕ 19 min 44 s VRAM 78G	✕ 44 min 58 s VRAM 83G
1920 × 1088 2.0MP · above native	✕ 19 min 46 s VRAM 78G	✕ 51 min 13 s VRAM 84G	✕ 1.4 h VRAM 88G	✕ 3.7 h VRAM 95G

Times are per whole clip (model load and text encoding included); the first run is slower due to kernel compilation. ⚠ = weight offloading / noticeably slower; ✕ = not recommended.

Reading the numbers

Every tier spills to the pagefile with the recommended combo — lower the resolution or step down a quant tier first. 28 high-end combinations are not recommended (pagefile pressure or >90 min per clip) — see the ✕ marks below. Already top-tier: stack 4–8 step distillation with Sage/Kitchen attention; for very long clips try SLA / VSA sparse attention.

Recommended combo (Research tier (96GB+ / multi-GPU))

Main model: minimax_h3_fl2va_bf16.safetensors（66.28GB）
Text encoder: qwen3vl_32b_minimax_h3_bf16.safetensors（51.51GB）
Acceleration: Official baseline (no acceleration) · ~1× sampling
Launch flags: defaults are fine

For research/training; not recommended for consumer use

Full-precision combination is 123.6GB; for production/multi-GPU serving see deployment paths such as vLLM

FAQ

Can the RTX PRO 6000 Blackwell 96GB run MiniMax H3?

Not with the default combo; smaller quantizations (GGUF Q3/Q2) at lower resolutions still work — see the model library and the calculator.

How fast is H3 on the RTX PRO 6000 Blackwell 96GB?

Try the 736×416 short-clip tier first.

How much RAM do I need?

At least undefinedGB of system RAM recommended (minimum 128GB plus a undefinedGB pagefile). With 32GB, always enable --fast-disk or weights spill into the pagefile.

How can I make it faster?

Stack step distillation (e.g. Official baseline (no acceleration)), enable SageAttention / Kitchen attention; for long clips try SLA / VSA sparse attention. All of it is switchable in step 2 of the calculator.

→ Run the numbers for your RAM and budget in the calculator (RTX PRO 6000 Blackwell 96GB pre-filled)

Same tier: —