DeepSeek V4 Pro Model Size: 1.6T Parameters, 49B Active, VRAM Needs

DeepSeek V4 Pro's official specs, unchanged between the April preview and the August 13 GA (0813) build^[4][5].

The 1.6T total parameter count is a measure of knowledge capacity; the 49B active count is what actually runs per token. V4 Pro activates only ~3% of its parameters per token, which is why inference stays practical despite the huge total^[4].

Contrast with a dense model: a dense 1.6T model would need 1.6T parameters of compute per token; V4 Pro needs 49B. That is the entire trick behind billion-scale MoE — near-dense knowledge at one-thirtieth the per-token compute.

DeepSeek's architecture details (CSA + HCA attention, mHC hyper-connections, Muon optimizer) are covered in the V4 Pro main guide; for this page the takeaway is simple: total size = capacity, active size = cost, context = memory.

• 1.6T total parameters (knowledge capacity)
• 49B active parameters per token (compute cost)
• ≈3% of the model activates per token
• Same 1M context / 384K output as V4 Flash, which is 284B total / 13B active

The FP4+FP8 weight mix puts the full model at a third-party estimate of roughly **862GB of VRAM** — the practical minimum for hosting the complete 0813 weights^[5].

DeepSeek's official vLLM recipe uses a 4xGB300 single node with DSpark speculative decoding enabled — the reference deployment for the GA build^[4].

One 80GB GPU (RTX 4090/5090-class) is only enough for distillation or small experiments; production serving needs a multi-GPU node. The hosted API is the realistic default for most teams.

# Rough VRAM estimate for V4 Pro (FP4 experts + FP8 rest)
weights ≈ 862 GB total

# GPU configurations that fit:
#  8 × H100/H200 80GB  = 640GB  → too small for full model
# 10 × H200 96GB       = 960GB  → fits with margin
#  4 × GB300           = fits  (DeepSeek's official vLLM reference node)

# With DSpark speculative decoding:
vllm serve deepseek-ai/DeepSeek-V4-Pro-0813 \
  --speculative-config '{"method":"dspark", ...}' \
  --tensor-parallel-size 8

Given the VRAM floor, the deployment ladder runs from API to self-hosted cluster^[2][4].

• Hosted API (recommended): deepseek-v4-pro at api.deepseek.com — $0.66/$1.32 input (miss), $1.98/$3.96 output (off-peak/peak), cache hits from $0.022
• Third-party routers: OpenRouter lists deepseek/deepseek-v4-pro-0813 at $0.435/$0.87 (pre-Aug-16 flat pricing shown on the router)
• Self-hosted: vLLM + DSpark on 4xGB300 (official reference) or a comparable multi-GPU node
• Distillation: single-GPU experiments on the full weights are not feasible; use the Flash weights (284B/13B) for edge experiments instead

The family split is 5.6x on total parameters and 3.8x on active parameters^[4].

The size gap maps to a capability gap and a price gap: Flash activates 13B per token and costs about a third of Pro on every line. DeepSeek's own 22-benchmark aggregation puts Flash at roughly 83-88% of Pro's quality — the reason the Pro vs Flash decision is a routing question, not a religious one.

For hardware planning: if the ~862GB Pro footprint is a non-starter, the Flash footprint (and the Flash model size page) is the pragmatic open-weight option.