Alibaba Qwen3.5

Qwen3.5-397B-A17B hardware requirements

Needs at least 4 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 32 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $16.00/hr.

Parameters

403B

Active per token

17.0B

mixture of experts

KV cache per token

30 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3.5-397B-A17B

The strongest part each vendor offers for Qwen3.5-397B-A17B 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.

8 x GB200 NVL72 GPU 186GB 4 x Instinct MI355X OAM 8 x Ascend 950DT 16 x Gaudi 2 96GB 8 x TPU v7 192GB

Memory fills at: 8 x GB200 NVL72 GPU 186GB 520 users, 4 x Instinct MI355X OAM 252 users, 8 x Ascend 950DT 249 users, 16 x Gaudi 2 96GB 552 users, 8 x TPU v7 192GB 559 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

4 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 621
Per hour: $23.96
Per million tokens: $10.72
Memory used: 73%

Rent on Daytona

Cheapest per hour

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 462
Per hour: $8.00
Per million tokens: $4.81
Memory used: 55%

Rent on RunPod

Cheapest per million tokens

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 859
Per hour: $8.00
Per million tokens: $2.59
Memory used: 58%

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; 32 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	4	51 %	3.2k tok/s	148 tok/s	—	—	—	Specs →
2	Rubin SXM	4	76 %	2.9k tok/s	140 tok/s	—	—	—	Specs →
3	TPU 8t announced	8	51 %	~2.1k tok/s	78.2 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	76 %	~1.3k tok/s	54.5 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	76 %	1.2k tok/s	50.7 tok/s	$23.96	$5.41	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	76 %	1.2k tok/s	50.7 tok/s	$49.68	$11.24	—	Rent →
7	GB200 NVL72 GPU 186GB	8	59 %	2.3k tok/s	95.8 tok/s	$84.00	$10.06	—	Rent →
8	Instinct MI350X OAM	4	76 %	1.2k tok/s	50.7 tok/s	$24.64	$5.69	4.0 kW	Rent →
9	TPU v7 192GB	8	57 %	2.1k tok/s	88.4 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	4	51 %	~2.3k tok/s	148 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	84 %	1.2k tok/s	50.7 tok/s	$19.96	$4.64	5.6 kW	Rent →
12	B200 SXM 180GB	8	61 %	2.2k tok/s	92.2 tok/s	$32.72	$4.14	8.0 kW	Rent →
13	Ascend 970 announced	4	76 %	1.7k tok/s	91.3 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	8	57 %	2.1k tok/s	95.8 tok/s	—	—	5.6 kW	Specs →
15	Jalapeño announced	8	51 %	~3.1k tok/s	184 tok/s	—	—	5.6 kW	Specs →
16	Instinct MI325X OAM	4	86 %	833 tok/s	38.0 tok/s	$15.20	$5.07	4.0 kW	Rent →
17	Instinct MI300X 192GB	8	57 %	1.4k tok/s	63.5 tok/s	$19.12	$3.68	6.0 kW	Rent →
18	BR100 2 nodes	16	87 %	935 tok/s	33.8 tok/s	—	—	8.8 kW	Specs →
19	Ascend 960 announced	4	76 %	991 tok/s	60.9 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	8	78 %	1.2k tok/s	57.5 tok/s	$19.84	$4.46	5.6 kW	Rent →
21	H100 SXM5 80GB 2 nodes	16	69 %	1.7k tok/s	70.8 tok/s	$29.28	$4.81	11.2 kW	Rent →
22	Data Center GPU Max 1550 128GB	8	86 %	902 tok/s	39.3 tok/s	—	—	4.8 kW	Specs →
23	H200 NVL 141GB	8	78 %	~1.2k tok/s	57.5 tok/s	$28.91	$6.89	4.8 kW	Rent →
24	H100 NVL 94GB 2 nodes	16	59 %	~1.8k tok/s	82.4 tok/s	$51.04	$7.90	6.4 kW	Rent →
25	Ascend 910C	8	86 %	875 tok/s	38.3 tok/s	—	—	4.8 kW	Specs →
26	H100 PCIe 80GB 2 nodes	16	69 %	1.1k tok/s	42.3 tok/s	$40.00	$10.40	5.6 kW	Rent →
27	H800 PCIe 80GB 2 nodes	16	69 %	1.1k tok/s	42.3 tok/s	—	—	5.6 kW	Specs →
28	Ascend 950DT	8	76 %	883 tok/s	47.9 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition 2 nodes	16	58 %	893 tok/s	37.9 tok/s	$28.80	$8.96	9.6 kW	Rent →
30	Ascend 950PR	8	86 %	464 tok/s	19.2 tok/s	—	—	7.2 kW	Specs →
31	MTT S5000 2 nodes	16	69 %	~819 tok/s	33.8 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition 2 nodes	16	58 %	818 tok/s	33.8 tok/s	$11.04	$3.75	9.6 kW	Rent →
33	Gaudi 3 128GB	8	86 %	1.2k tok/s	44.3 tok/s	—	—	7.2 kW	Specs →
34	TPU v5p 95GB 2 nodes	16	58 %	1.2k tok/s	58.4 tok/s	—	—	7.2 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition 2 nodes	16	58 %	859 tok/s	37.9 tok/s	$8.00	$2.59	4.8 kW	Rent →
36	Cloud AI 100 Ultra	8	86 %	~186 tok/s	6.6 tok/s	—	—	1.2 kW	Specs →
37	Gaudi 2 96GB 2 nodes	16	58 %	1.1k tok/s	51.8 tok/s	$17.60	$4.63	9.6 kW	Rent →
38	Atlas 350	8	99 %	403 tok/s	16.8 tok/s	—	—	4.8 kW	Specs →
39	Ascend 910B 2 nodes	16	87 %	772 tok/s	33.8 tok/s	—	—	6.4 kW	Specs →
40	MI250X 128GB	8	86 %	675 tok/s	39.3 tok/s	—	—	4.5 kW	Specs →
41	Instinct MI250 128GB	8	86 %	657 tok/s	39.3 tok/s	—	—	4.0 kW	Specs →
42	A800 PCIe 80GB 2 nodes	16	69 %	806 tok/s	40.9 tok/s	—	—	4.8 kW	Specs →
43	A800 SXM4 2 nodes	16	69 %	832 tok/s	43.1 tok/s	—	—	6.4 kW	Specs →
44	A100 PCIe 80GB 2 nodes	16	69 %	806 tok/s	40.9 tok/s	$21.60	$7.44	4.8 kW	Rent →
45	A100 SXM4 80GB 2 nodes	16	69 %	832 tok/s	43.1 tok/s	$22.40	$7.48	6.4 kW	Rent →
46	RTX PRO 5000 Blackwell 72GB 2 nodes	16	77 %	~596 tok/s	28.4 tok/s	—	—	4.8 kW	Specs →
47	Instinct MI210 PCIe 2 nodes	16	87 %	580 tok/s	34.6 tok/s	—	—	4.8 kW	Specs →
48	H20 141GB HBM3e	8	78 %	412 tok/s	47.9 tok/s	—	—	3.2 kW	Specs →
49	H20 96GB 2 nodes	16	58 %	727 tok/s	84.6 tok/s	—	—	6.4 kW	Specs →
50	Atlas 300I Duo 96GB 2 nodes	16	58 %	213 tok/s	8.6 tok/s	—	—	2.4 kW	Specs →
51	ATOM-Max 2 nodes	16	87 %	384 tok/s	21.6 tok/s	—	—	5.6 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 859 tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 1.4k tokens/s for $19.12/hr. Per rental dollar NVIDIA delivers 1.42x the tokens of AMD here.

NVIDIA 22 parts fit

Best value: 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 859 tok/s for $8.00/hr

Fewest GPUs: 4 x Rubin SXM, 2.9k tok/s

AMD 8 parts fit

Best value: 8 x Instinct MI300X 192GB, 1.4k tok/s for $19.12/hr

Fewest GPUs: 4 x Instinct MI455X, 3.2k tok/s

Huawei 6 parts fit

Fewest GPUs: 8 x Ascend 950DT, 883 tok/s

Intel 3 parts fit

Best value: 16 x Gaudi 2 96GB, 1.1k tok/s for $17.60/hr

Fewest GPUs: 8 x Gaudi 3 128GB, 1.2k tok/s

Google 2 parts fit

Fewest GPUs: 8 x TPU v7 192GB, 2.1k tok/s

Other 4 parts fit

Fewest GPUs: 8 x Cloud AI 100 Ultra, 186 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	125 x TP8	0	155k tok/s	$2,480	700 kW	$4.46
B	104 x GB200 NVL72 GPU 186GB	13 x TP8	0	30k tok/s	$1,092	—	$10.06

Size Qwen3.5-397B-A17B 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 Qwen3.5-397B-A17B need?

The weights take 807 GB at the native BF16 precision (807 GB at BF16, 403 GB at FP8, 227 GB at INT4). Each concurrent stream adds 30 KB of KV cache per token: 0.1 GB at 4,096 tokens and 1.0 GB at 32,768 tokens.

What is the bare minimum to run Qwen3.5-397B-A17B?

4 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00 per hour.

How many H100s do you need to run Qwen3.5-397B-A17B?

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

How much does Qwen3.5-397B-A17B cost per million tokens on an H100?

About $4.81 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 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.59.

What precision does Qwen3.5-397B-A17B 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 15 full-attention layers keep a per-token cache. The 45 linear-attention layers hold a fixed 189 MB state per sequence (64 value heads x 128 x 128 in FP32 per layer).