DeepSeek DeepSeek V4 Estimated geometry

DeepSeek-V4-Flash hardware requirements

Fits on one GB300 NVL72 GPU 288GB at INT4 / FP4. For 1,000 tokens/s at 32k context you need 4 x H100 SXM5 80GB at $7.32/hr or 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00/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 Model card

Top options for DeepSeek-V4-Flash

The strongest part each vendor offers for DeepSeek-V4-Flash at INT4 / FP4 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.

Tokens per second as users grow

Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.

1 x GB300 NVL72 GPU 288GB 1 x Instinct MI355X OAM 2 x Ascend 950DT 2 x Gaudi 2 96GB 1 x TPU v7 192GB 8 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 417 users, 1 x Instinct MI355X OAM 417 users, 2 x Ascend 950DT 413 users, 2 x Gaudi 2 96GB 72 users, 1 x TPU v7 192GB 76 users, 8 x Inferentia2 276 users. Roofline estimates at INT4 / FP4.

Bare minimum

Fewest GPUs

1 x GB300 NVL72 GPU 288GB

INT4 / FP4, 4,096 tokens, one stream

Tokens/s: 763
Per hour: $12.42
Per million tokens: $4.52
Memory used: 59%

Rent on DataCrunch

Cheapest per hour

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 322
Per hour: $1.00
Per million tokens: $0.86
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: 614
Per hour: $1.00
Per million tokens: $0.45
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; 67 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

#									Rent
1	Instinct MI455X	1	41 %	4.7k tok/s	179 tok/s	—	—	—	Specs →
2	Rubin SXM	1	62 %	4.4k tok/s	169 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	83 %	~1.4k tok/s	50.2 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	62 %	~1.6k tok/s	66.1 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	62 %	1.5k tok/s	61.5 tok/s	$5.99	$1.09	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	62 %	1.7k tok/s	61.5 tok/s	$12.42	$2.09	—	Rent →
7	GB200 NVL72 GPU 186GB	1	96 %	1.5k tok/s	61.5 tok/s	$10.50	$1.91	—	Rent →
8	Instinct MI350X OAM	1	62 %	1.5k tok/s	61.5 tok/s	$6.16	$1.14	1.0 kW	Rent →
9	TPU v7 192GB	1	93 %	~848 tok/s	56.7 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	41 %	~3.0k tok/s	179 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	68 %	1.6k tok/s	61.5 tok/s	$4.99	$0.85	1.4 kW	Rent →
12	B200 SXM 180GB	1	100 %	1.5k tok/s	59.2 tok/s	$4.09	$0.78	1.0 kW	Rent →
13	Ascend 970 announced	1	62 %	2.2k tok/s	111 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	93 %	1.4k tok/s	61.5 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	83 %	~2.7k tok/s	118 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	70 %	~560 tok/s	46.1 tok/s	$3.80	$1.89	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	93 %	~533 tok/s	40.7 tok/s	$2.39	$1.25	0.8 kW	Rent →
18	BR100	4	71 %	~910 tok/s	44.3 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	62 %	1.3k tok/s	73.8 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	64 %	~822 tok/s	70.1 tok/s	$4.96	$1.68	1.4 kW	Rent →
21	H100 SXM5 80GB	4	57 %	~1.3k tok/s	92.7 tok/s	$7.32	$1.51	2.8 kW	Rent →
22	TPU v6e 32GB	8	72 %	~1.7k tok/s	85.6 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	70 %	~640 tok/s	47.8 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	2	64 %	~736 tok/s	70.1 tok/s	$7.23	$2.73	1.2 kW	Rent →
25	H100 NVL 94GB	2	96 %	~684 tok/s	56.9 tok/s	$6.38	$2.59	0.8 kW	Rent →
26	Ascend 910C	2	70 %	~616 tok/s	46.7 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	4	57 %	~912 tok/s	55.3 tok/s	$10.00	$3.05	1.4 kW	Rent →
28	H800 PCIe 80GB	4	57 %	~912 tok/s	55.3 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	62 %	1.1k tok/s	58.4 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	94 %	636 tok/s	26.2 tok/s	$3.60	$1.57	1.2 kW	Rent →
31	Ascend 950PR	2	70 %	582 tok/s	23.4 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	4	57 %	~662 tok/s	44.3 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	94 %	581 tok/s	23.3 tok/s	$1.38	$0.66	1.2 kW	Rent →
34	Gaudi 3 128GB	2	70 %	~968 tok/s	54.0 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	95 %	~411 tok/s	40.4 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	94 %	614 tok/s	26.2 tok/s	$1.00	$0.45	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	70 %	~177 tok/s	8.0 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	94 %	~379 tok/s	35.8 tok/s	$2.20	$1.61	1.2 kW	Rent →
39	Atlas 350	2	80 %	496 tok/s	20.4 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	95 %	~532 tok/s	34.0 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	72 %	2.2k tok/s	93.6 tok/s	$9.52	$1.22	4.6 kW	Rent →
42	Ascend 910B	4	71 %	~584 tok/s	44.3 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	70 %	~378 tok/s	47.8 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	95 %	~438 tok/s	26.6 tok/s	$3.88	$2.46	1.2 kW	Rent →
45	Instinct MI250 128GB	2	70 %	~362 tok/s	47.8 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	95 %	~413 tok/s	23.9 tok/s	$3.04	$2.04	1.4 kW	Rent →
47	A800 PCIe 80GB	4	57 %	~534 tok/s	53.5 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	57 %	~542 tok/s	56.4 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	57 %	~534 tok/s	53.5 tok/s	$5.40	$2.81	1.2 kW	Rent →
50	A100 SXM4 80GB	4	57 %	~542 tok/s	56.4 tok/s	$5.60	$2.87	1.6 kW	Rent →
51	A100 PCIe 40GB	8	57 %	~937 tok/s	81.2 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	57 %	~937 tok/s	81.2 tok/s	$12.49	$3.70	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	97 %	~1.6k tok/s	70.2 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	95 %	~380 tok/s	26.6 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	72 %	~785 tok/s	62.7 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	72 %	~539 tok/s	30.1 tok/s	$6.64	$3.42	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	95 %	~813 tok/s	37.2 tok/s	$3.84	$1.31	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	63 %	~813 tok/s	37.2 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	72 %	986 tok/s	41.8 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	72 %	1.1k tok/s	46.8 tok/s	$5.76	$1.49	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	72 %	~488 tok/s	33.4 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	72 %	~540 tok/s	42.8 tok/s	—	—	—	Specs →
63	Trainium	8	72 %	~540 tok/s	42.8 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	72 %	~358 tok/s	64.2 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	72 %	~466 tok/s	31.8 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	95 %	~283 tok/s	23.9 tok/s	$3.28	$3.22	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	71 %	~343 tok/s	45.3 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	97 %	~531 tok/s	52.7 tok/s	$7.12	$3.73	3.6 kW	Rent →
69	A30 24GB	8	97 %	~517 tok/s	48.7 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	97 %	~366 tok/s	22.6 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	95 %	~246 tok/s	21.2 tok/s	$2.00	$2.26	1.2 kW	Rent →
72	A40 48GB	4	95 %	~232 tok/s	19.2 tok/s	$2.36	$2.83	1.2 kW	Rent →
73	H20 141GB HBM3e	2	64 %	~176 tok/s	58.4 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	2	94 %	~176 tok/s	58.4 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	97 %	~799 tok/s	35.1 tok/s	$4.56	$1.59	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	72 %	~513 tok/s	62.7 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	95 %	~177 tok/s	11.3 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	94 %	~93.5 tok/s	6.0 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	97 %	~427 tok/s	40.1 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	97 %	~375 tok/s	30.1 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	72 %	~462 tok/s	59.2 tok/s	$7.04	$4.24	2.0 kW	Rent →
82	ATOM-Max	4	71 %	~235 tok/s	28.3 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	97 %	~370 tok/s	31.3 tok/s	$8.81	$6.61	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	95 %	~218 tok/s	23.9 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	95 %	~218 tok/s	23.9 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	97 %	~424 tok/s	50.2 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	97 %	~266 tok/s	15.7 tok/s	$4.72	$4.92	0.6 kW	Rent →
88	L20 48GB	4	95 %	~214 tok/s	23.9 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	72 %	~324 tok/s	33.4 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	72 %	~318 tok/s	31.8 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	97 %	~288 tok/s	23.8 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	97 %	~238 tok/s	15.7 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	97 %	~516 tok/s	22.6 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	95 %	~166 tok/s	17.0 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	95 %	~174 tok/s	23.9 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	72 %	~294 tok/s	30.1 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	97 %	~275 tok/s	48.9 tok/s	$4.00	$4.04	2.8 kW	Rent →
98	A10G 24GB	8	97 %	~247 tok/s	31.3 tok/s	$8.05	$9.03	2.4 kW	Rent →
99	Atlas 300V Pro	4	95 %	~88.8 tok/s	5.7 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	97 %	~168 tok/s	10.7 tok/s	—	—	—	Specs →
101	Atlas 300V	8	97 %	~139 tok/s	10.7 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	97 %	~169 tok/s	52.7 tok/s	$3.68	$6.05	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	72 %	~140 tok/s	26.7 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	97 %	~118 tok/s	40.1 tok/s	$2.16	$5.07	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	83 %	~111 tok/s	23.4 tok/s	—	—	1.3 kW	Specs →

By vendor

At INT4 / FP4 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 614 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI355X OAM at about 1.5k tokens/s for $5.99/hr. Per rental dollar NVIDIA delivers 2.40x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 614 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 4.4k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI355X OAM, 1.5k tok/s for $5.99/hr

Fewest GPUs: 1 x Instinct MI455X, 4.7k tok/s

Huawei 10 parts fit

Fewest GPUs: 2 x Ascend 950DT, 1.1k tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 379 tok/s for $2.20/hr

Fewest GPUs: 2 x Gaudi 3 128GB, 968 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 848 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 540 tok/s

Other 7 parts fit

Fewest GPUs: 2 x Cloud AI 100 Ultra, 177 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 letter	Fleet	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A	1,000 x H200 SXM 141GB	500 x TP2	0	411k tok/s	$2,480	700 kW	$1.68
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	153k tok/s	$1,050	—	$1.91

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 GB300 NVL72 GPU 288GB 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 $1.51 per million output tokens at INT4 / FP4, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.45.

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.