DeepSeek DeepSeek V4 Estimated geometry

DeepSeek-V4-Flash hardware requirements

Fits on one Instinct MI350X OAM at INT4 / FP4. For 1,000 tokens/s at 32k context you need 4 x H100 SXM5 80GB at $7.16/hr or 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00/hr.

Parameters

291B

Active per token

13.8B

mixture of experts

KV cache per token

8.1 KB

16-bit cache

Native precision

FP4

Rent the cheapest fit: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition from $1.00/hr Model card

Bare minimum

Fewest GPUs

1 x Instinct MI350X OAM

INT4 / FP4, 4,096 tokens, one stream

Tokens/s: 770
Per hour: $6.16
Per million tokens: $2.22
Memory used: 59%

Rent on DigitalOcean

Cheapest per hour

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 4,096 tokens, one stream, live price

Tokens/s: 328
Per hour: $1.00
Per million tokens: $0.85
Memory used: 89%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 32,768 tokens, 32 streams

Tokens/s: 5.0k
Per hour: $1.00
Per million tokens: $0.06
Memory used: 94%

Rent on RunPod

Configurations

Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.

Precision Context Streams

Vendor

Priced only

Priced configurations only; 37 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

GPU	TP	Memory used	Tokens/s	Per stream	Per hour	Per M tokens	kW
Instinct MI455X OAM	1	41%	~34k	1.1k				Specs
Rubin SXM	1	62%	32k	1.0k				Specs
Instinct MI355X OAM	1	62%	12k	364			1.4	Specs
GB300 NVL72 GPU 288GB	1	62%	12k	364	$8.62	$0.21		Rent
GB200 NVL72 GPU 186GB	1	96%	12k	364	$10.50	$0.25		Rent
Instinct MI350X OAM	1	62%	12k	364	$6.16	$0.15	1.0	Rent
B300 SXM 262GB	1	68%	12k	364	$7.40	$0.18	1.4	Rent
B200 SXM 180GB	1	100%	11k	351	$4.09	$0.10	1.0	Rent
B100 SXM 192GB	1	93%	12k	364			0.7	Specs
Instinct MI325X OAM	1	70%	~8.7k	273	$3.80	$0.12	1.0	Rent
Instinct MI300X 192GB	1	93%	~7.7k	241	$2.39	$0.09	0.8	Rent
BR100	4	71%	~8.4k	262			2.2	Specs
H200 SXM 141GB	2	64%	~13k	415	$5.98	$0.12	1.4	Rent
H100 SXM5 80GB	4	57%	~18k	549	$7.16	$0.11	2.8	Rent
Data Center GPU Max 1550 128GB	2	70%	~9.1k	284			1.2	Specs
H200 NVL 141GB	2	64%	~13k	415	$7.58	$0.16	1.2	Rent
H100 NVL 94GB	2	96%	~11k	338	$6.38	$0.16	0.8	Rent
H100 PCIe 80GB	4	57%	~10k	328	$10.00	$0.26	1.4	Rent
H800 PCIe 80GB	4	57%	~10k	328			1.4	Specs
RTX PRO 6000 Blackwell Workstation Edition	2	94%	5.0k	155	$3.60	$0.20	1.2	Rent
RTX PRO 6000 Blackwell Server Edition	2	94%	~4.4k	138	$1.18	$0.07	1.2	Rent
RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	94%	5.0k	155	$1.00	$0.06	0.6	Rent
Data Center GPU Max 1100 48GB	4	95%	~6.4k	201			1.2	Specs
GeForce RTX 5090 32GB	8	72%	18k	555	$7.92	$0.12	4.6	Rent
MI250X 128GB	2	70%	~9.1k	284			1.1	Specs
RTX 6000 Ada 48GB	4	95%	~5.0k	157	$3.36	$0.19	1.2	Rent
Instinct MI250 128GB	2	70%	~9.1k	284			1.0	Specs
L40S 48GB	4	95%	~4.5k	142	$3.48	$0.21	1.4	Rent
A800 PCIe 80GB	4	57%	~10k	317			1.2	Specs
A100 PCIe 80GB	4	57%	~10k	317	$5.40	$0.15	1.2	Rent
A100 SXM4 80GB	4	57%	~11k	334	$5.60	$0.15	1.6	Rent
A100 PCIe 40GB	8	57%	~15k	482			2.0	Specs
A100 SXM4 40GB	8	57%	~15k	482	$10.32	$0.19	3.2	Rent
GeForce RTX 5090 D V2 24GB	8	97%	~13k	416			4.6	Specs
RTX 5880 Ada 48GB	4	95%	~5.0k	157			1.1	Specs
RTX 5000 Ada 32GB	8	72%	~5.7k	178	$6.64	$0.32	2.0	Rent
RTX PRO 5000 Blackwell 48GB	4	95%	~7.1k	220	$3.84	$0.15	1.2	Rent
RTX PRO 5000 Blackwell 72GB	4	63%	~7.1k	220			1.2	Specs
RTX PRO 4500 Blackwell 32GB	8	72%	~8.9k	278	$5.76	$0.18	1.6	Rent
Radeon AI PRO R9700 32GB	8	72%	~6.3k	198			2.4	Specs
Instinct MI100 32GB	8	72%	~12k	381			2.4	Specs
Arc Pro B70 32GB	8	72%	~6.0k	188			1.8	Specs
L40 48GB	4	95%	~4.5k	142	$3.28	$0.20	1.2	Rent
Instinct MI210 PCIe	4	71%	~8.6k	269			1.2	Specs
GeForce RTX 4090 24GB	8	97%	~10.0k	312	$4.80	$0.13	3.6	Rent
A30 24GB	8	97%	~9.2k	289	$5.86	$0.18	1.3	Rent
RTX 4500 Ada 24GB	8	97%	~4.3k	134			1.7	Specs
RTX A6000 48GB	4	95%	~4.0k	126	$2.00	$0.14	1.2	Rent
A40 48GB	4	95%	~3.7k	114	$1.96	$0.15	1.2	Rent
H20 141GB HBM3e	2	64%	~6.1k	191			0.8	Specs
H20 96GB	2	94%	~6.1k	191			0.8	Specs
RTX PRO 4000 Blackwell 24GB	8	97%	~6.7k	208	$4.56	$0.19	1.2	Rent
BI-V100 Tiangai 100	8	72%	~12k	372			2.0	Specs
RTX A5500 24GB	8	97%	~7.6k	238			1.8	Specs
V100S PCIe 32GB	8	72%	~11k	351	$7.04	$0.17	2.0	Rent
A10 24GB	8	97%	~5.9k	186	$10.32	$0.48	1.2	Rent
Radeon PRO W7900 Dual Slot 48GB	4	95%	~4.5k	142			1.2	Specs
Radeon PRO W7900 48GB	4	95%	~4.5k	142			1.2	Specs
Radeon RX 7900 XTX 24GB	8	97%	~9.5k	297			2.8	Specs
L4 24GB	8	97%	~3.0k	92.9	$3.92	$0.37	0.6	Rent
L20 48GB	4	95%	~4.5k	142			1.1	Specs
Arc Pro B65 32GB	8	72%	~6.0k	188			1.6	Specs
Arc Pro B60 24GB	8	97%	~4.5k	141			1.6	Specs
L2 24GB	8	97%	~3.0k	92.9				Specs
RTX PRO 4000 Blackwell SFF 24GB	8	97%	~4.3k	134			0.6	Specs
Radeon PRO W7800 48GB	4	95%	~4.5k	142			1.0	Specs
Radeon PRO W7800 32GB	8	72%	~5.7k	178			2.1	Specs
GeForce RTX 3090 24GB	8	97%	~9.3k	290	$4.00	$0.12	2.8	Rent
A10G 24GB	8	97%	~5.9k	186			2.4	Specs
GeForce RTX 3090 Ti 24GB	8	97%	~5.9k	185	$3.68	$0.17	3.6	Rent
Radeon PRO W6800 32GB	8	72%	~5.1k	159			2.0	Specs
RTX A5000 24GB	8	97%	~4.1k	128	$2.16	$0.15	1.8	Rent
Radeon PRO V710 28GB	8	83%	~4.1k	128			1.3	Specs

AMD vs NVIDIA

At INT4 / FP4 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 5.0k tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 7.7k tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.54x the tokens of AMD here.

NVIDIA 49 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 5.0k tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 32k tok/s

AMD 17 parts fit

Best value: 1 x Instinct MI300X 192GB, 7.7k tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X OAM, 34k tok/s

Intel 5 parts fit

Fewest GPUs: 2 x Data Center GPU Max 1550 128GB, 9.1k tok/s

Other 2 parts fit

Fewest GPUs: 4 x BR100, 8.4k tok/s

Fleet what-if

Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.

A GPUs Part

B GPUs Part

Fleet	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A 1,000 x H200 SXM 141GB	500 x TP2	0	6.65M	$2,990	700 kW	$0.12
B 100 x GB200 NVL72 GPU 186GB	100 x TP1	0	1.17M	$1,050	kW	$0.25

Size DeepSeek-V4-Flash for a tokens per second target

The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.

Open in the LLM GPU Calculator

Frequently asked questions

How much GPU memory does DeepSeek-V4-Flash need?

The weights take 164 GB at the native INT4 / FP4 precision (582 GB at BF16, 291 GB at FP8, 164 GB at INT4). Each concurrent stream adds 8.1 KB of KV cache per token: 0.0 GB at 4,096 tokens and 0.3 GB at 32,768 tokens.

What is the bare minimum to run DeepSeek-V4-Flash?

1 x Instinct MI350X OAM at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00 per hour.

How many H100s do you need to run DeepSeek-V4-Flash?

4 H100s at INT4 / FP4 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 4 H100s, and 1,000 tokens per second takes 4 H100s across 1 replicas.

How much does DeepSeek-V4-Flash cost per million tokens on an H100?

About $0.11 per million output tokens at INT4 / FP4, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.79 per GPU-hour. The cheapest part per token is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.06.

What precision does DeepSeek-V4-Flash ship in?

The published checkpoint is FP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.

All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. Experts are stored in FP4 and attention in FP8. The KV cache is a single 512-wide latent per layer compressed on most layers; 8.3 KB per token is an estimate from compress_ratios. Active parameters computed from the expert geometry.