Can DeepSeek V3 0324 run on the H100 SXM?
No. Even at 2-bit with 8K context, DeepSeek V3 0324 needs about 228 GB, more than the 76.0 GB usable on the H100 SXM.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 228 GB No | 228 GB No | 230 GB No | 236 GB No | — |
| 3-bit (Q3_K_M) • | 298 GB No | 298 GB No | 300 GB No | 307 GB No | — |
| 4-bit (Q4_K_M) • | 378 GB No | 378 GB No | 379 GB No | 386 GB No | — |
| 5-bit (Q5_K_M) • | 444 GB No | 444 GB No | 446 GB No | 452 GB No | — |
| 6-bit (Q6_K) • | 514 GB No | 514 GB No | 516 GB No | 522 GB No | — |
| 8-bit (Q8_0) • | 665 GB No | 665 GB No | 667 GB No | 674 GB No | — |
| 16-bit (F16/BF16) • | 1251 GB No | 1251 GB No | 1253 GB No | 1259 GB No | — |
• real GGUF file size from unsloth/DeepSeek-V3-0324-GGUF. Speed is a bandwidth ceiling (3350 GB/s), not a benchmark. Methodology.
Hardware that does fit it
More
DeepSeek V3 0324 on every GPU and Mac · Every model that runs on the H100 SXM · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (deepseek-ai/DeepSeek-V3-0324), NVIDIA specs. Data as of 2026-11-17.