VRAM Reference

DeepSeek V3 VRAM Requirements: FP16 / INT4 / KV-Cache (2026)

DeepSeek V3 runs on 671B parameters with a 128K-token context window. Weights are deterministic param math; KV-cache uses the registry layer geometry where published, otherwise a modeled GQA estimate — every figure is provenance-labeled below.

FP16 weights1340 GB

INT4 weights336 GB

Context128K tokens

Single-GPU verdictMulti-GPU (381 GB)

1340 GB FP16 weights

Calculated EstimateHIGH

SourceOpenGPU Radar models-registry.json (param-count weight model)

VerifiedSep 29, 2026

Value

1,340 GB

Methodology

weights = parametersB × bytes-per-parameter (FP16 = 2 B, INT4 = 0.5 B)

Refreshed daily from provider APIs and market scrapingSep 29, 2026

336 GB INT4 weights

Calculated EstimateHIGH

SourceOpenGPU Radar models-registry.json (param-count weight model)

VerifiedSep 29, 2026

Value

336 GB

Methodology

weights = parametersB × bytes-per-parameter (FP16 = 2 B, INT4 = 0.5 B)

Refreshed daily from provider APIs and market scrapingSep 29, 2026

128K context window

Manufacturer SpecHIGH

SourceModel vendor context-window specification

VerifiedSep 29, 2026

Value

128,000 tokens

Refreshed daily from provider APIs and market scrapingSep 29, 2026

Methodology →

How much VRAM does DeepSeek V3 need?

Weights: 1340 GB at FP16, 336 GB at INT4 (parameters × bytes-per-parameter: FP16 = 2 B, INT4 = 0.5 B).

Full stack at 128K: 1498.4 GB FP16 / 381.0 GB INT4 — including KV-cache, CUDA overhead, activations, and fragmentation headroom from the canonical VRAM engine.

Single-GPU verdict: Multi-GPU required — 381 GB INT4 exceeds single-GPU capacity; registry recommendation: 8x H200.

VRAM by Precision & Context

From the canonical VRAM engine: weights + KV-cache + CUDA overhead + activations + fragmentation headroom. KV-cache uses published layer geometry where available; otherwise a modeled GQA estimate (labeled in the badge).

Precision	Context	Weights	KV-Cache	Total (est.)
FP16	4K	1342.0 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1,342 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: FP16 = 2 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: fp16 bytesPerParam: 2 Refreshed daily from provider APIs and market scrapingSep 26, 2026	0.59 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 0.59 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 4096 batchSize: 1 quantization: fp16 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	1478.6 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1478.61 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: fp16 contextLength: 4096 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 2 bytesPerKvElement: 2 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026
FP16	33K	1342.0 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1,342 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: FP16 = 2 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: fp16 bytesPerParam: 2 Refreshed daily from provider APIs and market scrapingSep 26, 2026	4.75 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 4.75 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 32768 batchSize: 1 quantization: fp16 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	1483.2 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1483.18 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: fp16 contextLength: 32768 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 2 bytesPerKvElement: 2 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026
FP16	128K	1342.0 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1,342 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: FP16 = 2 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: fp16 bytesPerParam: 2 Refreshed daily from provider APIs and market scrapingSep 26, 2026	18.55 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 18.55 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 128000 batchSize: 1 quantization: fp16 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	1498.4 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 1498.37 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: fp16 contextLength: 128000 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 2 bytesPerKvElement: 2 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026
INT4	4K	335.5 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 335.5 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: INT4 = 0.5 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: int4 bytesPerParam: 0.5 Refreshed daily from provider APIs and market scrapingSep 26, 2026	0.30 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 0.3 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 4096 batchSize: 1 quantization: int4 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	371.1 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 371.14 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: int4 contextLength: 4096 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 0.5 bytesPerKvElement: 1 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026
INT4	33K	335.5 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 335.5 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: INT4 = 0.5 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: int4 bytesPerParam: 0.5 Refreshed daily from provider APIs and market scrapingSep 26, 2026	2.38 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 2.38 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 32768 batchSize: 1 quantization: int4 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	373.4 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 373.42 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: int4 contextLength: 32768 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 0.5 bytesPerKvElement: 1 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026
INT4	128K	335.5 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 335.5 GB Methodology Formula: parameterCountB × bytesPerParam. Precision: INT4 = 0.5 bytes/param. (parameterCountB in billions, result in GB). Assumptions & Parameters paramCountB: 671 quantization: int4 bytesPerParam: 0.5 Refreshed daily from provider APIs and market scrapingSep 26, 2026	9.28 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 9.28 GB Methodology GQA KV Cache: kvBytes = 2 × numLayers × numKvHeads × headDim × contextLength × batchSize × bytesPerKvElement. GQA ratio: 4:1. Assumptions & Parameters numLayers: 38 numKvHeads: 8 headDim: 128 contextLength: 128000 batchSize: 1 quantization: int4 gqaRatio: 4 Refreshed daily from provider APIs and market scrapingSep 26, 2026	381.0 GB Calculated EstimateHIGH SourceDeterministic VRAM Canonical Engine VerifiedSep 26, 2026 Value 381.01 GB Methodology Total = weights + kvCache + cudaOverhead + activation + fragmentationHeadroom (10%). Assumptions & Parameters paramCountB: 671 quantization: int4 contextLength: 128000 batchSize: 1 numLayers: 38 numKvHeads: 8 headDim: 128 gqaRatio: 4 architecture: deepseek-v3 bytesPerParam: 0.5 bytesPerKvElement: 1 cudaOverheadGb: 1.2 activationGb: 0.4 fragmentationHeadroomPct: 10 isMla: true Refreshed daily from provider APIs and market scrapingSep 26, 2026

Cloud GPUs That Fit DeepSeek V3

Fit = full-stack total (weights + KV at 128K) ≤ GPU VRAM. Rates are the lowest observed on-demand rows in data/providers.json, refreshed daily.

GPU	VRAM	FP16 fit	INT4 fit	Lowest on-demand
H200 SXM5	141 GB	OOM	OOM	$2.79/hr (Vast.ai)
B200 Blackwell	192 GB	OOM	OOM	$3.99/hr (Vast.ai)
H100 SXM5	80 GB	OOM	OOM	$1.89/hr (Vast.ai)
A100 80GB SXM4	80 GB	OOM	OOM	$1.59/hr (Lambda Labs)
L40S	48 GB	OOM	OOM	$0.69/hr (Vast.ai)
RTX 4090	24 GB	OOM	OOM	$0.34/hr (Vast.ai)

Recommended GPUs for DeepSeek V3

Minimum = smallest single GPU that fits the full INT4 stack (381.0 GB). Optimal = registry recommendation, falling back to the lowest observed on-demand rate among fitting cards. Every card links to its canonical GPU page.

No single GPU in the candidate set fits 381 GB INT4 — DeepSeek V3 needs multi-GPU or a smaller model. Size it in the calculator →

Workload Scenarios Using DeepSeek V3

Costed workload guides that price DeepSeek V3 against verified cloud rates and registry VRAM math.

Batch

Batch Inference

Offline batch inference economics: why the cheapest large-VRAM GPU leads when latency is free and preemption risk is priced in.

Next steps

⚡

Open in VRAM calculatorPre-filled with DeepSeek V3

✓

Full model profileBenchmarks, providers & run commands

GPU

Frequently Asked Questions

How much VRAM does DeepSeek V3 need?▾

DeepSeek V3 needs 1340 GB at FP16 (weights) and 336 GB at INT4 from the param-count weight model. Full-stack totals at 128K context are 1498.4 GB (FP16) and 381.0 GB (INT4) including KV-cache, CUDA overhead, and fragmentation headroom.

Can DeepSeek V3 run on a single GPU?▾

Yes — the registry recommendation is 8x H200, and the NVIDIA H200 SXM5 (141GB HBM3e) fits the FP16 workload single-GPU. FP16 full-stack needs 1498.4 GB.

What context length fits DeepSeek V3?▾

Registry context window is 128,000 tokens. KV-cache grows linearly with context — the quantization table shows totals at 4K, 33K, 128K context for FP16 and INT4.

Weight model: parameters × bytes-per-parameter. KV-cache: GQA formula or modeled estimate where layer geometry is unpublished.Methodology →