Alibaba Qwen3.5

Qwen3.5-122B-A10B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 4 x H100 SXM5 80GB at $7.32/hr or 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00/hr.

Parameters

125B

Active per token

10.0B

mixture of experts

KV cache per token

24 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3.5-122B-A10B

The strongest part each vendor offers for Qwen3.5-122B-A10B 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.

2 x GB200 NVL72 GPU 186GB 2 x Instinct MI325X OAM 4 x Ascend 910C 4 x Gaudi 2 96GB 2 x TPU v7 192GB 16 x Inferentia2

Memory fills at: 2 x GB200 NVL72 GPU 186GB 112 users, 2 x Instinct MI325X OAM 253 users, 4 x Ascend 910C 251 users, 4 x Gaudi 2 96GB 121 users, 2 x TPU v7 192GB 124 users, 16 x Inferentia2 237 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 292
Per hour: $5.99
Per million tokens: $5.69
Memory used: 90%

Rent on Daytona

Cheapest per hour

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 235
Per hour: $2.00
Per million tokens: $2.37
Memory used: 68%

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: 520
Per hour: $2.00
Per million tokens: $1.07
Memory used: 77%

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	67 %	1.8k tok/s	89.7 tok/s	—	—	—	Specs →
2	Rubin SXM	2	51 %	3.0k tok/s	161 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	68 %	~1.2k tok/s	47.8 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	51 %	~1.4k tok/s	62.9 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	51 %	1.3k tok/s	58.5 tok/s	$11.98	$2.51	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	51 %	1.3k tok/s	58.5 tok/s	$24.84	$5.21	—	Rent →
7	GB200 NVL72 GPU 186GB	2	79 %	1.3k tok/s	58.5 tok/s	$21.00	$4.40	—	Rent →
8	Instinct MI350X OAM	2	51 %	1.3k tok/s	58.5 tok/s	$12.32	$2.65	2.0 kW	Rent →
9	TPU v7 192GB	2	76 %	1.2k tok/s	54.0 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	67 %	~1.2k tok/s	89.7 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	56 %	1.3k tok/s	58.5 tok/s	$9.98	$2.16	2.8 kW	Rent →
12	B200 SXM 180GB	2	81 %	1.2k tok/s	56.3 tok/s	$8.18	$1.82	2.0 kW	Rent →
13	Ascend 970 announced	2	51 %	1.7k tok/s	105 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	76 %	1.2k tok/s	58.5 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	68 %	~1.7k tok/s	113 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	57 %	881 tok/s	43.9 tok/s	$7.60	$2.40	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	76 %	814 tok/s	38.8 tok/s	$4.78	$1.63	1.5 kW	Rent →
18	BR100	8	58 %	1.1k tok/s	41.9 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	51 %	1.0k tok/s	70.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	52 %	1.3k tok/s	66.5 tok/s	$9.92	$2.11	2.8 kW	Rent →
21	H100 SXM5 80GB	4	92 %	1.0k tok/s	46.4 tok/s	$7.32	$1.97	2.8 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	59 %	2.0k tok/s	75.7 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	57 %	966 tok/s	45.4 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	52 %	~1.2k tok/s	66.5 tok/s	$14.46	$3.29	2.4 kW	Rent →
25	H100 NVL 94GB	4	78 %	~1.1k tok/s	54.1 tok/s	$12.76	$3.29	1.6 kW	Rent →
26	Ascend 910C	4	57 %	935 tok/s	44.4 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	4	92 %	661 tok/s	27.7 tok/s	$10.00	$4.20	1.4 kW	Rent →
28	H800 PCIe 80GB	4	92 %	661 tok/s	27.7 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	4	51 %	912 tok/s	55.5 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	77 %	544 tok/s	24.8 tok/s	$7.20	$3.67	2.4 kW	Rent →
31	Ascend 950PR	4	57 %	502 tok/s	22.2 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	4	92 %	~502 tok/s	22.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	77 %	501 tok/s	22.1 tok/s	$2.76	$1.53	2.4 kW	Rent →
34	Gaudi 3 128GB	4	57 %	1.3k tok/s	51.3 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	77 %	689 tok/s	38.3 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	77 %	520 tok/s	24.8 tok/s	$2.00	$1.07	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	57 %	~209 tok/s	7.6 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	77 %	627 tok/s	34.0 tok/s	$4.40	$1.95	2.4 kW	Rent →
39	Atlas 350	4	66 %	435 tok/s	19.4 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	77 %	747 tok/s	32.2 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	59 %	1.3k tok/s	82.8 tok/s	$19.04	$4.20	9.2 kW	Rent →
42	Ascend 910B	8	58 %	884 tok/s	41.9 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	57 %	690 tok/s	45.4 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	77 %	~600 tok/s	25.1 tok/s	$7.76	$3.59	2.4 kW	Rent →
45	Instinct MI250 128GB	4	57 %	670 tok/s	45.4 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	77 %	553 tok/s	22.6 tok/s	$6.08	$3.05	2.8 kW	Rent →
47	A800 PCIe 80GB	4	92 %	476 tok/s	26.8 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	92 %	490 tok/s	28.3 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	92 %	476 tok/s	26.8 tok/s	$5.40	$3.15	1.2 kW	Rent →
50	A100 SXM4 80GB	4	92 %	490 tok/s	28.3 tok/s	$5.60	$3.17	1.6 kW	Rent →
51	A100 PCIe 40GB	8	93 %	792 tok/s	40.7 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	93 %	792 tok/s	40.7 tok/s	$12.49	$4.38	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	79 %	~1.3k tok/s	62.1 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	77 %	~555 tok/s	25.1 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	59 %	1.1k tok/s	55.5 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	59 %	~663 tok/s	26.6 tok/s	$13.28	$5.57	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	77 %	~673 tok/s	35.2 tok/s	$7.68	$3.17	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	8	51 %	~673 tok/s	35.2 tok/s	—	—	2.4 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	59 %	784 tok/s	37.0 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	59 %	843 tok/s	41.4 tok/s	$11.52	$3.79	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	59 %	660 tok/s	29.6 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	59 %	778 tok/s	37.9 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	59 %	778 tok/s	37.9 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	59 %	667 tok/s	56.8 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	59 %	~629 tok/s	28.1 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	77 %	448 tok/s	22.6 tok/s	$6.56	$4.07	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	58 %	632 tok/s	42.9 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	79 %	833 tok/s	46.6 tok/s	$14.24	$4.75	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	79 %	797 tok/s	43.1 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	79 %	~472 tok/s	20.0 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	77 %	392 tok/s	20.1 tok/s	$4.00	$2.83	2.4 kW	Rent →
72	A40 48GB	8	77 %	364 tok/s	18.2 tok/s	$4.72	$3.60	2.4 kW	Rent →
73	H20 141GB HBM3e	4	52 %	395 tok/s	55.5 tok/s	—	—	1.6 kW	Specs →
74	H20 96GB	4	77 %	395 tok/s	55.5 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	79 %	~625 tok/s	31.1 tok/s	$9.12	$4.05	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	59 %	863 tok/s	55.5 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	77 %	248 tok/s	10.7 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	77 %	131 tok/s	5.7 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	79 %	~657 tok/s	35.5 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	79 %	543 tok/s	26.6 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	59 %	788 tok/s	52.4 tok/s	$14.08	$4.96	4.0 kW	Rent →
82	ATOM-Max	8	58 %	422 tok/s	26.8 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	79 %	547 tok/s	27.7 tok/s	$17.63	$8.95	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	77 %	379 tok/s	22.6 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	77 %	379 tok/s	22.6 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	79 %	706 tok/s	44.4 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	79 %	336 tok/s	13.9 tok/s	$9.44	$7.81	1.2 kW	Rent →
88	L20 48GB	8	77 %	374 tok/s	22.6 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	59 %	515 tok/s	29.6 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	59 %	~500 tok/s	28.1 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	79 %	~422 tok/s	21.1 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	79 %	316 tok/s	13.9 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	79 %	~404 tok/s	20.0 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	77 %	282 tok/s	16.1 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	77 %	324 tok/s	22.6 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	59 %	467 tok/s	26.6 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	79 %	512 tok/s	43.3 tok/s	$8.00	$4.34	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	79 %	421 tok/s	27.7 tok/s	$16.10	$10.63	4.8 kW	Rent →
99	Atlas 300V Pro	8	77 %	124 tok/s	5.4 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	79 %	220 tok/s	9.5 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	79 %	198 tok/s	9.5 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	79 %	350 tok/s	46.6 tok/s	$7.36	$5.84	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	59 %	265 tok/s	23.7 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	79 %	248 tok/s	35.5 tok/s	$4.32	$4.83	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	68 %	214 tok/s	20.7 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 520 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 814 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, 520 tok/s for $2.00/hr

Fewest GPUs: 2 x Rubin SXM, 3.0k tok/s

AMD 18 parts fit

Best value: 2 x Instinct MI300X 192GB, 814 tok/s for $4.78/hr

Fewest GPUs: 1 x Instinct MI455X, 1.8k tok/s

Huawei 10 parts fit

Fewest GPUs: 4 x Ascend 910C, 935 tok/s

Intel 7 parts fit

Best value: 4 x Gaudi 2 96GB, 627 tok/s for $4.40/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 1.3k tok/s

Google 4 parts fit

Fewest GPUs: 2 x TPU v7 192GB, 1.2k tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 778 tok/s

Other 7 parts fit

Fewest GPUs: 4 x MTT S5000, 502 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	327k tok/s	$2,480	700 kW	$2.11
B	100 x GB200 NVL72 GPU 186GB	50 x TP2	0	66k tok/s	$1,050	—	$4.40

Size Qwen3.5-122B-A10B 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 Qwen3.5-122B-A10B need?

The weights take 250 GB at the native BF16 precision (250 GB at BF16, 125 GB at FP8, 70.4 GB at INT4). Each concurrent stream adds 24 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.8 GB at 32,768 tokens.

What is the bare minimum to run Qwen3.5-122B-A10B?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00 per hour.

How many H100s do you need to run Qwen3.5-122B-A10B?

4 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 4 H100s, and 1,000 tokens per second takes 4 H100s across 1 replicas.

How much does Qwen3.5-122B-A10B cost per million tokens on an H100?

About $1.97 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 $1.07.

What precision does Qwen3.5-122B-A10B 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 12 full-attention layers keep a per-token cache. The 36 linear-attention layers hold a fixed 151 MB state per sequence.