Can DeepSeek V4 Flash Vision Exp run on the H100 SXM?
Not fully in VRAM. Even at 2-bit with 8K context, DeepSeek V4 Flash Vision Exp needs about 111 GB, more than the 76.0 GB usable on the H100 SXM. It can still run with offload (slower): Q4_K_M with part of the model in system RAM, about 8.7–18 tokens/sec (estimate). Details and command.
Runs with offload (slower)
Runs with offload 4-bit (Q4_K_M) at 8K context: all attention layers plus the experts of 18 of 46 MoE layers on the H100 SXM (about 71.5 GB), and the experts of the other 28 layers in system RAM (about 103 GB). Needs a PC with at least 128 GB of system RAM. Estimated speed: 8.7–18 tokens/sec (usable for chat), versus a bandwidth ceiling of ~360 tok/s if it fit fully in VRAM.
llama-server -m deepseek-v4-flash-vision-exp-Q4_K_M.gguf -c 8192 -ngl 999 --n-cpu-moe 28
Estimate, not a benchmark. The GGUF repo we found (ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF) has no Q4_K_M file; point -m at your own Q4_K_M GGUF. The speed range spans dual-channel DDR4-3200 to DDR5-6000 system RAM; your CPU, RAM speed and llama.cpp version move it. If llama.cpp runs out of GPU memory, raise --n-cpu-moe by 1-2; if VRAM is left over, lower it for more speed. Recent llama.cpp builds can also pick the split automatically. How we estimate offload.
+ 0.5 GB with image input. The figures on this page are for text only. With image input enabled, the vision projector (mmproj) file loads alongside the model. llama.cpp's -hf downloads and loads it automatically unless you add --no-mmproj. Calculator with image input on.
Newer compressed-attention design: long-context figures are an upper bound.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 110 GB No | 111 GB No | 113 GB No | 122 GB No | — |
| 3-bit (Q3_K_M) | 140 GB No | 140 GB No | 142 GB No | 151 GB No | — |
| 4-bit (Q4_K_M) | 174 GB No | 175 GB No | 177 GB No | 186 GB No | — |
| 5-bit (Q5_K_M) | 203 GB No | 204 GB No | 206 GB No | 215 GB No | — |
| 6-bit (Q6_K) | 234 GB No | 234 GB No | 236 GB No | 245 GB No | — |
| 8-bit (Q8_0) | 302 GB No | 303 GB No | 305 GB No | 314 GB No | — |
| 16-bit (F16/BF16) | 568 GB No | 569 GB No | 571 GB No | 580 GB No | — |
• real GGUF file size from ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF. Speed is a bandwidth ceiling (3350 GB/s), not a benchmark. Methodology.
Hardware that does fit it
H200Mac M2 Ultra 192GBMac M3 Ultra 256GBMac M3 Ultra 512GB
What fits instead
- No listed GPU with a list price runs DeepSeek V4 Flash Vision Exp fully in memory. The smallest listed hardware that does is the H200 (141 GB).
- Largest popular model that fits the H100 SXM: gpt-oss 120B (117B, MoE) at F16, about 61.9 GB at 8K context.
"Runs" and "fits" mean the whole model in GPU memory at 8K context (4-bit or better for the model pick). Prices are US launch MSRPs, not today's street prices. "Popular" means the 60 most-downloaded models we track.
More
DeepSeek V4 Flash Vision Exp 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
Guides: Best GPU for local LLMs by budget · Can I run gpt-oss, Llama 70B or Qwen3? · Mac vs PC for local LLMs
Data: Hugging Face Hub (deepseek-ai/DeepSeek-V4-Flash-Vision-Exp), NVIDIA specs. Model data last checked 2026-11-18; refreshed hourly.