Alibaba Qwen3

Qwen3-32B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 20 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $10.00/hr.

Parameters

32.8B

Active per token

All

dense model

KV cache per token

256 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3-32B

The strongest part each vendor offers for Qwen3-32B at BF16 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 1 x Ascend 950DT 1 x Gaudi 3 128GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 24 users, 1 x Instinct MI355X OAM 24 users, 1 x Ascend 950DT 8 users, 1 x Gaudi 3 128GB 6 users, 1 x TPU v7 192GB 13 users, 4 x Inferentia2 6 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 88.9
Per hour: $5.99
Per million tokens: $18.72
Memory used: 24%

Rent on Daytona

Cheapest per hour

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 4,096 tokens, one stream, live price

Tokens/s: 19.8
Per hour: $0.50
Per million tokens: $7.00
Memory used: 73%

Rent on RunPod

Cheapest per million tokens

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 229
Per hour: $2.00
Per million tokens: $2.43
Memory used: 93%

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; 68 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	82 %	774 tok/s	51.3 tok/s	—	—	—	Specs →
2	Rubin SXM	2	61 %	1.3k tok/s	92.1 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	82 %	~582 tok/s	27.3 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	61 %	~631 tok/s	36.0 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	61 %	605 tok/s	33.5 tok/s	$11.98	$5.50	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	61 %	603 tok/s	33.5 tok/s	$24.84	$11.44	—	Rent →
7	GB200 NVL72 GPU 186GB	2	95 %	603 tok/s	33.5 tok/s	$21.00	$9.67	—	Rent →
8	Instinct MI350X OAM	2	61 %	582 tok/s	33.5 tok/s	$12.32	$5.88	2.0 kW	Rent →
9	TPU v7 192GB	2	92 %	557 tok/s	30.9 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	82 %	~475 tok/s	51.3 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	67 %	575 tok/s	33.5 tok/s	$9.98	$4.82	2.8 kW	Rent →
12	B200 SXM 180GB	2	98 %	564 tok/s	32.2 tok/s	$8.18	$4.03	2.0 kW	Rent →
13	Ascend 970 announced	2	61 %	702 tok/s	60.3 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	92 %	508 tok/s	33.5 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	82 %	~645 tok/s	64.5 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	69 %	380 tok/s	25.1 tok/s	$7.60	$5.55	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	92 %	358 tok/s	22.2 tok/s	$4.78	$3.71	1.5 kW	Rent →
18	BR100	8	70 %	556 tok/s	24.0 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	61 %	386 tok/s	40.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	63 %	560 tok/s	38.1 tok/s	$9.92	$4.92	2.8 kW	Rent →
21	H100 SXM5 80GB	8	56 %	882 tok/s	50.2 tok/s	$14.64	$4.61	5.6 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	71 %	967 tok/s	43.3 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	69 %	427 tok/s	26.0 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	63 %	~509 tok/s	38.1 tok/s	$14.46	$7.89	2.4 kW	Rent →
25	H100 NVL 94GB	4	95 %	~464 tok/s	30.9 tok/s	$12.76	$7.63	1.6 kW	Rent →
26	Ascend 910C	4	69 %	412 tok/s	25.4 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	8	56 %	584 tok/s	30.0 tok/s	$20.00	$9.51	2.8 kW	Rent →
28	H800 PCIe 80GB	8	56 %	584 tok/s	30.0 tok/s	—	—	2.8 kW	Specs →
29	Ascend 950DT	4	62 %	366 tok/s	31.7 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	93 %	244 tok/s	14.2 tok/s	$7.20	$8.19	2.4 kW	Rent →
31	Ascend 950PR	4	69 %	229 tok/s	12.7 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	8	56 %	~432 tok/s	24.0 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	93 %	228 tok/s	12.7 tok/s	$2.76	$3.36	2.4 kW	Rent →
34	Gaudi 3 128GB	4	69 %	612 tok/s	29.3 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	94 %	285 tok/s	21.9 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	93 %	229 tok/s	14.2 tok/s	$2.00	$2.43	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	69 %	~107 tok/s	4.3 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	93 %	262 tok/s	19.4 tok/s	—	—	2.4 kW	Specs →
39	Atlas 350	4	79 %	198 tok/s	11.1 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	94 %	344 tok/s	18.4 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	71 %	493 tok/s	47.4 tok/s	$19.04	$10.73	9.2 kW	Rent →
42	Ascend 910B	8	70 %	389 tok/s	24.0 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	69 %	270 tok/s	26.0 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	94 %	~281 tok/s	14.4 tok/s	$7.92	$7.84	2.4 kW	Rent →
45	Instinct MI250 128GB	4	69 %	260 tok/s	26.0 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	94 %	262 tok/s	12.9 tok/s	$6.08	$6.44	2.8 kW	Rent →
47	A800 PCIe 80GB	8	56 %	370 tok/s	29.0 tok/s	—	—	2.4 kW	Specs →
48	A800 SXM4	8	56 %	378 tok/s	30.5 tok/s	—	—	3.2 kW	Specs →
49	A100 PCIe 80GB	8	56 %	370 tok/s	29.0 tok/s	$10.80	$8.10	2.4 kW	Rent →
50	A100 SXM4 80GB	8	56 %	378 tok/s	30.5 tok/s	$10.80	$7.93	3.2 kW	Rent →
51	A100 PCIe 40GB 2 nodes	16	57 %	596 tok/s	41.1 tok/s	—	—	4.0 kW	Specs →
52	A100 SXM4 40GB 2 nodes	16	57 %	596 tok/s	41.1 tok/s	$24.98	$11.65	6.4 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	96 %	~542 tok/s	35.5 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	94 %	~250 tok/s	14.4 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	71 %	493 tok/s	31.7 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	71 %	~317 tok/s	15.2 tok/s	$13.28	$11.62	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	94 %	~285 tok/s	20.1 tok/s	$7.68	$7.49	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	8	62 %	~285 tok/s	20.1 tok/s	—	—	2.4 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	71 %	346 tok/s	21.2 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	71 %	366 tok/s	23.7 tok/s	$11.52	$8.74	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	71 %	299 tok/s	16.9 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	71 %	339 tok/s	21.7 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	71 %	339 tok/s	21.7 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	71 %	246 tok/s	32.5 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	71 %	~285 tok/s	16.1 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	94 %	192 tok/s	12.9 tok/s	$6.56	$9.49	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	70 %	246 tok/s	24.5 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	96 %	344 tok/s	26.7 tok/s	$6.40	$5.17	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	96 %	333 tok/s	24.7 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	96 %	~220 tok/s	11.4 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	94 %	167 tok/s	11.5 tok/s	$4.00	$6.65	2.4 kW	Rent →
72	A40 48GB	8	94 %	157 tok/s	10.4 tok/s	$4.72	$8.36	2.4 kW	Rent →
73	H20 141GB HBM3e	4	63 %	134 tok/s	31.7 tok/s	—	—	1.6 kW	Specs →
74	H20 96GB	4	93 %	134 tok/s	31.7 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	96 %	~270 tok/s	17.8 tok/s	$9.12	$9.39	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	71 %	341 tok/s	31.7 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	94 %	115 tok/s	6.1 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	93 %	60.7 tok/s	3.2 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	96 %	~275 tok/s	20.3 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	96 %	236 tok/s	15.2 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	71 %	308 tok/s	30.0 tok/s	$14.08	$12.70	4.0 kW	Rent →
82	ATOM-Max	8	70 %	167 tok/s	15.3 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	96 %	235 tok/s	15.9 tok/s	$17.63	$20.87	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	94 %	153 tok/s	12.9 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	94 %	153 tok/s	12.9 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	96 %	280 tok/s	25.4 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	96 %	159 tok/s	7.9 tok/s	$9.44	$16.54	1.2 kW	Rent →
88	L20 48GB	8	94 %	150 tok/s	12.9 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	71 %	211 tok/s	16.9 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	71 %	~206 tok/s	16.1 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	96 %	~182 tok/s	12.1 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	96 %	145 tok/s	7.9 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	96 %	~175 tok/s	11.4 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	94 %	115 tok/s	9.2 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	94 %	125 tok/s	12.9 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	71 %	191 tok/s	15.2 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	96 %	189 tok/s	24.7 tok/s	$8.00	$11.78	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	96 %	165 tok/s	15.9 tok/s	$16.10	$27.12	4.8 kW	Rent →
99	Atlas 300V Pro	8	94 %	57.4 tok/s	3.1 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	96 %	101 tok/s	5.4 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	96 %	86.9 tok/s	5.4 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	96 %	120 tok/s	26.7 tok/s	$7.36	$17.05	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	71 %	96.7 tok/s	13.5 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	96 %	84.4 tok/s	20.3 tok/s	$4.32	$14.22	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	82 %	76.9 tok/s	11.8 tok/s	—	—	2.5 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 229 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 358 tokens/s for $4.78/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 229 tok/s for $2.00/hr

Fewest GPUs: 2 x Rubin SXM, 1.3k tok/s

AMD 18 parts fit

Best value: 2 x Instinct MI300X 192GB, 358 tok/s for $4.78/hr

Fewest GPUs: 1 x Instinct MI455X, 774 tok/s

Huawei 10 parts fit

Fewest GPUs: 4 x Ascend 910C, 412 tok/s

Intel 7 parts fit

Fewest GPUs: 4 x Gaudi 3 128GB, 612 tok/s

Google 4 parts fit

Fewest GPUs: 2 x TPU v7 192GB, 557 tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 339 tok/s

Other 7 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 107 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	250 x TP4	0	140k tok/s	$2,480	700 kW	$4.92
B	100 x GB200 NVL72 GPU 186GB	50 x TP2	0	30k tok/s	$1,050	—	$9.67

Size Qwen3-32B 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 Qwen3-32B need?

The weights take 65.5 GB at the native BF16 precision (65.5 GB at BF16, 32.8 GB at FP8, 18.4 GB at INT4). Each concurrent stream adds 256 KB of KV cache per token: 1.1 GB at 4,096 tokens and 8.6 GB at 32,768 tokens.

What is the bare minimum to run Qwen3-32B?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.50 per hour.

How many H100s do you need to run Qwen3-32B?

one H100 at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 8 H100s, and 1,000 tokens per second takes 16 H100s across 2 replicas.

How much does Qwen3-32B cost per million tokens on an H100?

About $4.61 per million output tokens at BF16, 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 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.43.

What precision does Qwen3-32B ship in?

The published checkpoint is BF16. FP8 and INT4 figures describe post-training quantisations that halve and quarter the footprint at a small accuracy cost.

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.