Architecture truth card — DeepSeek-V4-Pro-0813
Exact safetensors params1,650,497,936,906 (1650.5B)
Hidden layers61
KV heads (GQA)1
Head dim512
Data snapshot (UTC)2026-11-18T18:29:13+00:00
Snapshot SHA-256d89025de2d15…
← 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 | 3074.29 / 0.95 / 155.21 Total 3230.46 GiB | 3074.29 / 3.81 / 155.21 Total 3233.32 GiB | 3074.29 / 15.25 / 155.21 Total 3244.76 GiB |
| FP8/INT8 8-bit | 1537.15 / 0.95 / 78.36 Total 1616.46 GiB | 1537.15 / 3.81 / 78.36 Total 1619.32 GiB | 1537.15 / 15.25 / 78.36 Total 1630.75 GiB |
| INT4/AWQ 4.5-bit | 864.64 / 0.95 / 44.73 Total 910.33 GiB | 864.64 / 3.81 / 44.73 Total 913.19 GiB | 864.64 / 15.25 / 44.73 Total 924.63 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-pro-0813 --ctx 32768
Live cheapest GPU configurations
| Scenario | Cheapest fitting rig (top 3) | Total VRAM | Lowest $/hr | 730 h/mo | Deploy |
|---|---|---|---|---|---|
| INT4 @ 32K ctx | 8x H200 NVL 143GB | 1144 GiB | $4.00 | $2920.00 | Deploy on RunPodCheck Vast.ai Verified |
| INT4 @ 32K ctx | 8x MI300X 192GB | 1536 GiB | $4.00 | $2920.00 | Deploy on RunPodCheck Vast.ai Verified |
| INT4 @ 32K ctx | 4x B300 288GB | 1152 GiB | $27.76 | $20264.80 | Deploy on RunPodCheck Vast.ai Verified |
| FP8 @ 32K ctx | 8x B300 288GB | 2304 GiB | $55.52 | $40529.60 | Deploy on RunPodCheck Vast.ai Verified |
What over-provisioning costs you
DeepSeek-V4-Pro-0813 at INT4 / 32K context fits (8x H200 NVL 143GB) for $2920.00/mo (730 h); FP8 needs the next rig up (8x B300 288GB) at $40529.60/mo. The $40529.60 instead of $2920.00 monthly gap is the price of skipping 4-bit quantization.