aliteq.

How much VRAM DeepSeek V4 Pro (MoE) needs, and what it costs to run

Some links on this page are referral links: we may earn a commission, at no cost to you.

Trying to decide what to buy? The best GPU for DeepSeek V4 Pro (MoE) →

1598.8B paramsMoE · 384 experts, 6 active1024k max contextmit

Just want to use it? Per-token API

$1.60in·$3.20out

per 1M tokens · DeepSeek V4 Pro on Novita AI · read Oct 29, 2026

No GPU to manage: you pay per token, which suits occasional use. Your prompts run on Novita's servers, not your own machine.

Price as listed on Novita's pricing page on Oct 29, 2026; it may have changed.

Memory needed, by quantisation

At 16k context. Weights + KV cache + ~0.8 GB overhead. DeepSeek V4 Pro (MoE) ships in FP4 + FP8, so its own release leads and no bigger copy is listed: it would only store the same numbers in more bytes. Memory is in binary GB (1,024³ bytes), the unit GPU memory is sold in; Hugging Face and Ollama list downloads in decimal GB, about 7% more.

QuantWeightsKV cacheTotalQuality
FP4 + FP8official805.3 GB0.1 GB806.2 GBThe model's own release (FP4 + FP8, a 864.7 GB download): the best quality there is. A bigger copy only stores the same numbers in more bytes. Its 4-bit GGUF builds keep the 4-bit experts as they are, so they come out about this size too.

How these numbers are produced

  • Model shape comes from the model's own config.json — 61 layers, 128 attention heads, 1 KV heads.
  • Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
  • KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
  • Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
  • This is a mixture-of-experts model. Generation reads only the experts active for each token (6 of 384), so throughput is scored on ~49.0B active parameters, not the full 1599B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Compressed attention (CSA + HCA, model card): each layer caches one shared 512-dim KV entry per compressed block, 30 layers at 1/4 of tokens (CSA) and 31 at 1/128 (HCA) per compress_ratios in config.json, plus a fixed 128-token sliding window and CSA indexer keys (128-dim at 1/4). That is ~4,900 cached numbers per token, stored mostly as FP8 in DeepSeek's implementation (the same math for V4-Flash matches the ~3.6 KB/token implied by the V4.1-Flash card's "1/4 of V4-Flash" = 890 B), so ~5.1 KB per token. Mapped so the engine's fp16 formula (2 x layers x kv_heads x head_dim x 2 B) gives that: 61 layers x 1 KV head x head_dim 21 = 5,124 B/token (~5 GB at 1M context). head_dim 21 is a mapping, not the model's 512; a q8/q4 KV setting in the calculator would understate it. Vendor: 1.6T total / 49B active; params = HF safetensors total; experts FP4, rest FP8. HF config.json + model card read 29 Oct 2026.
  • Rental prices are pulled hourly from provider APIs (last updated Thu, 12 Nov 2026 22:20:09 GMT). Spot/interruptible pricing can change or vanish without notice.

Try it interactively

Drag the context slider and watch the KV cache fill the card

Worth reading before you buy

Other models