Architecture truth card — DeepSeek-V4-Flash-DSpark
← Open the interactive VRAM & cost calculator on the homepage
Exact VRAM matrix — 3 precisions × 3 context lengths
| Precision | 8192 tokens ctx | 32768 tokens ctx | 131072 tokens ctx |
|---|---|---|---|
| FP16/BF16 16-bit | 566.58 / 0.67 / 29.83 Total 597.08 GiB | 566.58 / 2.69 / 29.83 Total 599.10 GiB | 566.58 / 10.75 / 29.83 Total 607.16 GiB |
| FP8/INT8 8-bit | 283.29 / 0.67 / 15.66 Total 299.63 GiB | 283.29 / 2.69 / 15.66 Total 301.64 GiB | 283.29 / 10.75 / 15.66 Total 309.70 GiB |
| INT4/AWQ 4.5-bit | 159.35 / 0.67 / 9.47 Total 169.49 GiB | 159.35 / 2.69 / 9.47 Total 171.51 GiB | 159.35 / 10.75 / 9.47 Total 179.57 GiB |
Each cell: Weights GiB / GQA KV Cache GiB / Runtime Overhead GiB; bold line = Total VRAM GiB. Concurrency = 1, FP16 KV cache (2 bytes), overhead = 1.5 + 0.05 × weights GiB.
Terminal speed-math — no install needed
Re-run this exact sizing from your own terminal:
curl -fsSL https://gpuvram.pages.dev/vram-cli.py | python3 - deepseek-v4-flash-dspark --ctx 32768
Live cheapest GPU configurations
| Scenario | Cheapest fitting rig (top 3) | Total VRAM | Lowest $/hr | 730 h/mo | Deploy |
|---|---|---|---|---|---|
| INT4 @ 32K ctx | 1x MI300X 192GB | 192 GiB | $0.50 | $365.00 | Deploy on RunPodCheck Vast.ai Verified |
| INT4 @ 32K ctx | 2x H200 NVL 143GB | 286 GiB | $1.00 | $730.00 | Deploy on RunPodCheck Vast.ai Verified |
| INT4 @ 32K ctx | 2x MI300X 192GB | 384 GiB | $1.00 | $730.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP8 @ 32K ctx | 2x MI300X 192GB | 384 GiB | $1.00 | $730.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP8 @ 32K ctx | 4x H200 NVL 143GB | 572 GiB | $2.00 | $1460.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP8 @ 32K ctx | 4x MI300X 192GB | 768 GiB | $2.00 | $1460.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP16 @ 32K ctx | 4x MI300X 192GB | 768 GiB | $2.00 | $1460.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP16 @ 32K ctx | 8x H200 NVL 143GB | 1144 GiB | $4.00 | $2920.00 | Deploy on RunPodCheck Vast.ai Verified |
| FP16 @ 32K ctx | 8x MI300X 192GB | 1536 GiB | $4.00 | $2920.00 | Deploy on RunPodCheck Vast.ai Verified |
What over-provisioning costs you
Running DeepSeek-V4-Flash-DSpark production traffic at INT4 / 32K context on the cheapest fitting rig (1x MI300X 192GB) bills $365.00 per month (730 h). Defaulting to full FP16 forces a larger rig — 4x MI300X 192GB at $1460.00/mo. Paying $1460.00 instead of $365.00 every single month is the price of skipping 4-bit AWQ quantization; if your quality budget allows INT4, that gap is pure waste.