Alibaba Qwen3.5

Qwen3.5-27B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 6 x H100 SXM5 80GB at $10.98/hr or 12 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $6.00/hr.

Parameters

27.8B

Active per token

All

dense model

KV cache per token

64 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3.5-27B

The strongest part each vendor offers for Qwen3.5-27B 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 910C 1 x Gaudi 2 96GB 1 x TPU v7 192GB 2 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 96 users, 1 x Instinct MI355X OAM 96 users, 1 x Ascend 910C 29 users, 1 x Gaudi 2 96GB 15 users, 1 x TPU v7 192GB 56 users, 2 x Inferentia2 1 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 106
Per hour: $5.99
Per million tokens: $15.74
Memory used: 20%

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: 23.6
Per hour: $0.50
Per million tokens: $5.89
Memory used: 61%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 197
Per hour: $1.00
Per million tokens: $1.41
Memory used: 70%

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	31 %	1.2k tok/s	135 tok/s	—	—	—	Specs →
2	Rubin SXM	1	47 %	1.0k tok/s	128 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	62 %	~571 tok/s	37.9 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	47 %	~562 tok/s	50.0 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	47 %	546 tok/s	46.5 tok/s	$5.99	$3.05	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	47 %	544 tok/s	46.5 tok/s	$12.42	$6.35	—	Rent →
7	GB200 NVL72 GPU 186GB	1	72 %	544 tok/s	46.5 tok/s	$10.50	$5.36	—	Rent →
8	Instinct MI350X OAM	1	47 %	517 tok/s	46.5 tok/s	$6.16	$3.31	1.0 kW	Rent →
9	TPU v7 192GB	1	70 %	502 tok/s	42.9 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	31 %	~667 tok/s	135 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	51 %	508 tok/s	46.5 tok/s	$4.99	$2.73	1.4 kW	Rent →
12	B200 SXM 180GB	1	75 %	501 tok/s	44.7 tok/s	$4.09	$2.27	1.0 kW	Rent →
13	Ascend 970 announced	1	47 %	546 tok/s	83.6 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	70 %	427 tok/s	46.5 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	62 %	~484 tok/s	89.5 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	52 %	320 tok/s	34.9 tok/s	$3.80	$3.30	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	70 %	308 tok/s	30.8 tok/s	$2.39	$2.16	0.8 kW	Rent →
18	BR100	4	53 %	580 tok/s	33.5 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	47 %	288 tok/s	55.8 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	96 %	246 tok/s	27.9 tok/s	$2.48	$2.80	0.7 kW	Rent →
21	H100 SXM5 80GB	2	85 %	417 tok/s	37.0 tok/s	$3.66	$2.44	1.4 kW	Rent →
22	TPU v6e 32GB	8	54 %	1.1k tok/s	64.7 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	53 %	371 tok/s	36.2 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	1	96 %	~217 tok/s	27.9 tok/s	$3.61	$4.63	0.6 kW	Rent →
25	H100 NVL 94GB	2	72 %	~390 tok/s	43.0 tok/s	$6.38	$4.54	0.8 kW	Rent →
26	Ascend 910C	2	53 %	357 tok/s	35.3 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	2	85 %	290 tok/s	22.1 tok/s	$5.00	$4.78	0.7 kW	Rent →
28	H800 PCIe 80GB	2	85 %	290 tok/s	22.1 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	94 %	150 tok/s	23.2 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	70 %	216 tok/s	19.8 tok/s	$3.60	$4.63	1.2 kW	Rent →
31	Ascend 950PR	2	53 %	207 tok/s	17.7 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	2	85 %	~207 tok/s	17.7 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	70 %	206 tok/s	17.6 tok/s	$1.38	$1.86	1.2 kW	Rent →
34	Gaudi 3 128GB	2	53 %	595 tok/s	40.8 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	71 %	229 tok/s	30.5 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	70 %	197 tok/s	19.8 tok/s	$1.00	$1.41	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	53 %	~115 tok/s	6.0 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	70 %	212 tok/s	27.0 tok/s	$2.20	$2.88	1.2 kW	Rent →
39	Atlas 350	2	60 %	177 tok/s	15.5 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	71 %	318 tok/s	25.7 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	54 %	402 tok/s	70.8 tok/s	$9.52	$6.58	4.6 kW	Rent →
42	Ascend 910B	4	53 %	338 tok/s	33.5 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	53 %	205 tok/s	36.2 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	71 %	~264 tok/s	20.1 tok/s	$3.88	$4.08	1.2 kW	Rent →
45	Instinct MI250 128GB	2	53 %	196 tok/s	36.2 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	71 %	252 tok/s	18.1 tok/s	$3.04	$3.35	1.4 kW	Rent →
47	A800 PCIe 80GB	2	85 %	157 tok/s	21.4 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	85 %	158 tok/s	22.5 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	85 %	157 tok/s	21.4 tok/s	$2.70	$4.79	0.6 kW	Rent →
50	A100 SXM4 80GB	2	85 %	158 tok/s	22.5 tok/s	$2.80	$4.91	0.8 kW	Rent →
51	A100 PCIe 40GB	4	86 %	281 tok/s	32.5 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	86 %	281 tok/s	32.5 tok/s	$6.24	$6.17	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	73 %	~492 tok/s	53.1 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	71 %	~223 tok/s	20.1 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	54 %	451 tok/s	47.4 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	54 %	~332 tok/s	22.8 tok/s	$6.64	$5.56	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	71 %	~235 tok/s	28.1 tok/s	$3.84	$4.54	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	94 %	~124 tok/s	14.8 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	54 %	322 tok/s	31.6 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	54 %	334 tok/s	35.4 tok/s	$5.76	$4.79	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	54 %	287 tok/s	25.3 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	54 %	310 tok/s	32.4 tok/s	—	—	—	Specs →
63	Trainium	8	54 %	310 tok/s	32.4 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	54 %	189 tok/s	48.5 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	54 %	~275 tok/s	24.0 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	71 %	161 tok/s	18.1 tok/s	$3.28	$5.65	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	53 %	186 tok/s	34.3 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	73 %	296 tok/s	39.8 tok/s	$7.12	$6.69	3.6 kW	Rent →
69	A30 24GB	8	73 %	290 tok/s	36.9 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	73 %	~220 tok/s	17.1 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	71 %	139 tok/s	16.1 tok/s	$2.00	$3.99	1.2 kW	Rent →
72	A40 48GB	4	71 %	132 tok/s	14.6 tok/s	$2.36	$4.96	1.2 kW	Rent →
73	H20 141GB HBM3e	1	96 %	47.5 tok/s	23.2 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	2	70 %	90.2 tok/s	44.1 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	73 %	~244 tok/s	26.5 tok/s	$4.56	$5.18	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	54 %	280 tok/s	47.4 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	71 %	106 tok/s	8.5 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	70 %	55.8 tok/s	4.5 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	73 %	~240 tok/s	30.3 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	73 %	215 tok/s	22.8 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	54 %	250 tok/s	44.8 tok/s	$7.04	$7.82	2.0 kW	Rent →
82	ATOM-Max	4	53 %	128 tok/s	21.4 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	73 %	211 tok/s	23.7 tok/s	$8.81	$11.63	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	71 %	120 tok/s	18.1 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	71 %	120 tok/s	18.1 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	73 %	232 tok/s	37.9 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	73 %	162 tok/s	11.8 tok/s	$4.72	$8.11	0.6 kW	Rent →
88	L20 48GB	4	71 %	118 tok/s	18.1 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	54 %	180 tok/s	25.3 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	54 %	~177 tok/s	24.0 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	73 %	~164 tok/s	18.0 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	73 %	141 tok/s	11.8 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	73 %	~159 tok/s	17.1 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	71 %	92.0 tok/s	12.8 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	71 %	93.5 tok/s	18.1 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	54 %	163 tok/s	22.8 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	73 %	145 tok/s	37.0 tok/s	$4.00	$7.65	2.8 kW	Rent →
98	A10G 24GB	8	73 %	134 tok/s	23.7 tok/s	$8.05	$16.65	2.4 kW	Rent →
99	Atlas 300V Pro	4	71 %	53.0 tok/s	4.3 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	73 %	100 tok/s	8.1 tok/s	—	—	—	Specs →
101	Atlas 300V	8	73 %	80.4 tok/s	8.1 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	73 %	86.9 tok/s	39.8 tok/s	$3.68	$11.77	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	54 %	73.7 tok/s	20.2 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	73 %	60.7 tok/s	30.3 tok/s	$2.16	$9.88	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	62 %	57.9 tok/s	17.7 tok/s	—	—	1.3 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 197 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 308 tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 197 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 1.0k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 308 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 1.2k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 150 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 212 tok/s for $2.20/hr

Fewest GPUs: 2 x Gaudi 3 128GB, 595 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 502 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 310 tok/s

Other 7 parts fit

Fewest GPUs: 2 x MTT S5000, 207 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	1,000 x TP1	0	246k tok/s	$2,480	700 kW	$2.80
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	54k tok/s	$1,050	—	$5.36

Size Qwen3.5-27B 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 Qwen3.5-27B need?

The weights take 55.6 GB at the native BF16 precision (55.6 GB at BF16, 27.8 GB at FP8, 15.6 GB at INT4). Each concurrent stream adds 64 KB of KV cache per token: 0.3 GB at 4,096 tokens and 2.1 GB at 32,768 tokens.

What is the bare minimum to run Qwen3.5-27B?

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 Qwen3.5-27B?

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 2 H100s, and 1,000 tokens per second takes 6 H100s across 3 replicas.

How much does Qwen3.5-27B cost per million tokens on an H100?

About $2.44 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 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.41.

What precision does Qwen3.5-27B 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. Only the 16 full-attention layers keep a per-token cache. The 48 linear-attention layers hold a fixed 151 MB state per sequence (48 value heads x 128 x 128 in FP32 per layer).