Alibaba Qwen3.6

Qwen3.6-35B-A3B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 2 x H100 SXM5 80GB at $3.66/hr or 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00/hr.

Parameters

36.0B

Active per token

3.0B

mixture of experts

KV cache per token

20 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3.6-35B-A3B

The strongest part each vendor offers for Qwen3.6-35B-A3B 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 2 96GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 279 users, 1 x Instinct MI355X OAM 279 users, 1 x Ascend 950DT 90 users, 1 x Gaudi 2 96GB 26 users, 1 x TPU v7 192GB 153 users, 4 x Inferentia2 64 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 963
Per hour: $5.99
Per million tokens: $1.73
Memory used: 26%

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: 215
Per hour: $0.50
Per million tokens: $0.65
Memory used: 79%

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: 804
Per hour: $1.00
Per million tokens: $0.35
Memory used: 52%

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	23 %	5.3k tok/s	247 tok/s	—	—	—	Specs →
2	Rubin SXM	1	34 %	4.7k tok/s	233 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	46 %	~1.8k tok/s	69.2 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	34 %	~2.1k tok/s	91.2 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	34 %	2.0k tok/s	84.8 tok/s	$5.99	$0.82	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	34 %	2.0k tok/s	84.8 tok/s	$12.42	$1.71	—	Rent →
7	GB200 NVL72 GPU 186GB	1	53 %	2.0k tok/s	84.8 tok/s	$10.50	$1.44	—	Rent →
8	Instinct MI350X OAM	1	34 %	2.0k tok/s	84.8 tok/s	$6.16	$0.86	1.0 kW	Rent →
9	TPU v7 192GB	1	52 %	1.9k tok/s	78.2 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	23 %	~3.8k tok/s	247 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	38 %	2.0k tok/s	84.8 tok/s	$4.99	$0.70	1.4 kW	Rent →
12	B200 SXM 180GB	1	55 %	1.9k tok/s	81.6 tok/s	$4.09	$0.59	1.0 kW	Rent →
13	Ascend 970 announced	1	34 %	2.8k tok/s	153 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	52 %	1.8k tok/s	84.8 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	46 %	~2.7k tok/s	163 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	39 %	1.4k tok/s	63.6 tok/s	$3.80	$0.77	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	52 %	1.3k tok/s	56.2 tok/s	$2.39	$0.53	0.8 kW	Rent →
18	BR100	2	78 %	884 tok/s	32.2 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	34 %	1.6k tok/s	102 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	71 %	1.1k tok/s	50.9 tok/s	$2.48	$0.64	0.7 kW	Rent →
21	H100 SXM5 80GB	2	63 %	1.6k tok/s	67.5 tok/s	$3.66	$0.64	1.4 kW	Rent →
22	TPU v6e 32GB	4	80 %	1.7k tok/s	62.5 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	78 %	784 tok/s	34.7 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	71 %	~1.0k tok/s	50.9 tok/s	$3.61	$1.00	0.6 kW	Rent →
25	H100 NVL 94GB	2	53 %	~1.7k tok/s	78.5 tok/s	$6.38	$1.06	0.8 kW	Rent →
26	Ascend 910C	1	78 %	760 tok/s	33.9 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	2	63 %	1.0k tok/s	40.3 tok/s	$5.00	$1.38	0.7 kW	Rent →
28	H800 PCIe 80GB	2	63 %	1.0k tok/s	40.3 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	69 %	762 tok/s	42.4 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	52 %	837 tok/s	36.1 tok/s	$3.60	$1.20	1.2 kW	Rent →
31	Ascend 950PR	1	78 %	404 tok/s	17.0 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	2	63 %	~768 tok/s	32.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	52 %	767 tok/s	32.2 tok/s	$1.38	$0.50	1.2 kW	Rent →
34	Gaudi 3 128GB	1	78 %	1.0k tok/s	39.2 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	2	53 %	1.1k tok/s	55.7 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	52 %	804 tok/s	36.1 tok/s	$1.00	$0.35	0.6 kW	Rent →
37	Cloud AI 100 Ultra	1	78 %	~164 tok/s	5.8 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	2	52 %	983 tok/s	49.3 tok/s	$2.20	$0.62	1.2 kW	Rent →
39	Atlas 350	1	89 %	351 tok/s	14.8 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	53 %	1.1k tok/s	46.9 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	4	80 %	1.1k tok/s	68.4 tok/s	$4.76	$1.16	2.3 kW	Rent →
42	Ascend 910B	2	78 %	722 tok/s	32.2 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	78 %	581 tok/s	34.7 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	4	53 %	~915 tok/s	36.6 tok/s	$3.88	$1.18	1.2 kW	Rent →
45	Instinct MI250 128GB	1	78 %	565 tok/s	34.7 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	4	53 %	841 tok/s	33.0 tok/s	$3.04	$1.00	1.4 kW	Rent →
47	A800 PCIe 80GB	2	63 %	750 tok/s	39.0 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	63 %	774 tok/s	41.1 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	63 %	750 tok/s	39.0 tok/s	$2.70	$1.00	0.6 kW	Rent →
50	A100 SXM4 80GB	2	63 %	774 tok/s	41.1 tok/s	$2.80	$1.01	0.8 kW	Rent →
51	A100 PCIe 40GB	4	63 %	1.2k tok/s	59.3 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	63 %	1.2k tok/s	59.3 tok/s	$6.24	$1.40	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	54 %	~2.1k tok/s	96.9 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	53 %	~856 tok/s	36.6 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	4	80 %	999 tok/s	45.8 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	80 %	~569 tok/s	22.0 tok/s	$3.32	$1.62	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	53 %	~1.1k tok/s	51.3 tok/s	$3.84	$1.01	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	70 %	~556 tok/s	27.1 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	80 %	687 tok/s	30.5 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	80 %	743 tok/s	34.2 tok/s	$2.88	$1.08	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	80 %	575 tok/s	24.4 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	80 %	685 tok/s	31.3 tok/s	—	—	—	Specs →
63	Trainium	4	80 %	685 tok/s	31.3 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	80 %	619 tok/s	46.9 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	80 %	~548 tok/s	23.2 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	4	53 %	699 tok/s	33.0 tok/s	$3.28	$1.30	1.2 kW	Rent →
67	Instinct MI210 PCIe	2	78 %	537 tok/s	33.0 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	8	54 %	1.4k tok/s	72.6 tok/s	$7.12	$1.41	3.6 kW	Rent →
69	A30 24GB	8	54 %	1.3k tok/s	67.2 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	54 %	~772 tok/s	31.1 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	53 %	613 tok/s	29.3 tok/s	$2.00	$0.91	1.2 kW	Rent →
72	A40 48GB	4	53 %	568 tok/s	26.6 tok/s	$2.36	$1.15	1.2 kW	Rent →
73	H20 141GB HBM3e	1	71 %	349 tok/s	42.4 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	2	52 %	663 tok/s	80.6 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	54 %	~1.0k tok/s	48.4 tok/s	$4.56	$1.22	1.2 kW	Rent →
76	BI-V100 Tiangai 100	4	80 %	783 tok/s	45.8 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	53 %	380 tok/s	15.6 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	52 %	200 tok/s	8.2 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	54 %	~1.1k tok/s	55.4 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	54 %	904 tok/s	41.5 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	4	80 %	717 tok/s	43.3 tok/s	$3.52	$1.36	1.0 kW	Rent →
82	ATOM-Max	2	78 %	356 tok/s	20.6 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	8	54 %	915 tok/s	43.2 tok/s	$8.81	$2.68	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	53 %	603 tok/s	33.0 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	53 %	603 tok/s	33.0 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	54 %	1.2k tok/s	69.2 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	54 %	547 tok/s	21.6 tok/s	$4.72	$2.40	0.6 kW	Rent →
88	L20 48GB	4	53 %	595 tok/s	33.0 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	80 %	462 tok/s	24.4 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	80 %	~447 tok/s	23.2 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	8	54 %	~704 tok/s	32.9 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	54 %	519 tok/s	21.6 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	54 %	~673 tok/s	31.1 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	53 %	446 tok/s	23.4 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	53 %	522 tok/s	33.0 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	4	80 %	418 tok/s	22.0 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	8	54 %	896 tok/s	67.5 tok/s	$4.00	$1.24	2.8 kW	Rent →
98	A10G 24GB	8	54 %	722 tok/s	43.2 tok/s	$8.05	$3.09	2.4 kW	Rent →
99	Atlas 300V Pro	4	53 %	191 tok/s	7.8 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	54 %	360 tok/s	14.8 tok/s	—	—	—	Specs →
101	Atlas 300V	8	54 %	329 tok/s	14.8 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	54 %	630 tok/s	72.6 tok/s	$3.68	$1.62	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	4	80 %	247 tok/s	19.5 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	8	54 %	448 tok/s	55.4 tok/s	$2.16	$1.34	1.8 kW	Rent →
105	Radeon PRO V710 28GB	4	92 %	200 tok/s	17.1 tok/s	—	—	0.6 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 804 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 1.3k 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, 804 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 4.7k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 1.3k tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 5.3k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 762 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 983 tok/s for $2.20/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 1.0k tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 1.9k tok/s

AWS 2 parts fit

Fewest GPUs: 4 x Inferentia2, 685 tok/s

Other 7 parts fit

Fewest GPUs: 1 x Cloud AI 100 Ultra, 164 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	1.07M tok/s	$2,480	700 kW	$0.64
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	202k tok/s	$1,050	—	$1.44

Size Qwen3.6-35B-A3B 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 Qwen3.6-35B-A3B need?

The weights take 71.9 GB at the native BF16 precision (71.9 GB at BF16, 36.0 GB at FP8, 20.2 GB at INT4). Each concurrent stream adds 20 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.7 GB at 32,768 tokens.

What is the bare minimum to run Qwen3.6-35B-A3B?

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 Qwen3.6-35B-A3B?

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 2 H100s across 1 replicas.

How much does Qwen3.6-35B-A3B cost per million tokens on an H100?

About $0.64 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 $0.35.

What precision does Qwen3.6-35B-A3B 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 10 of 40 layers keep a per-token cache; the 30 linear-attention layers hold a fixed 63 MB of FP32 state per sequence. Active parameters per the Qwen model card.