Alibaba Qwen3.8 Estimated geometry

Qwen3.8-Flash-Next hardware requirements

Needs at least 2 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.64/hr or 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00/hr.

Parameters

180B

Active per token

6.0B

mixture of experts

KV cache per token

24 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3.8-Flash-Next

The strongest part each vendor offers for Qwen3.8-Flash-Next 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.

4 x GB200 NVL72 GPU 186GB 2 x Instinct MI355X OAM 4 x Ascend 950DT 4 x Gaudi 2 96GB 2 x TPU v7 192GB 16 x Inferentia2

Memory fills at: 4 x GB200 NVL72 GPU 186GB 386 users, 2 x Instinct MI355X OAM 211 users, 4 x Ascend 950DT 209 users, 4 x Gaudi 2 96GB 7 users, 2 x TPU v7 192GB 9 users, 16 x Inferentia2 128 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

2 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 921
Per hour: $11.98
Per million tokens: $3.61
Memory used: 65%

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: 389
Per hour: $2.00
Per million tokens: $1.43
Memory used: 98%

Rent on RunPod

Cheapest per million tokens

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 1.1k
Per hour: $4.00
Per million tokens: $1.00
Memory used: 53%

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; 56 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	94 %	2.1k tok/s	87.1 tok/s	—	—	—	Specs →
2	Rubin SXM	2	70 %	3.5k tok/s	156 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	94 %	~1.3k tok/s	46.4 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	70 %	~1.6k tok/s	61.1 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	70 %	1.5k tok/s	56.8 tok/s	$11.98	$2.27	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	70 %	1.5k tok/s	56.8 tok/s	$24.84	$4.71	—	Rent →
7	GB200 NVL72 GPU 186GB	4	54 %	2.8k tok/s	108 tok/s	$42.00	$4.20	—	Rent →
8	Instinct MI350X OAM	2	70 %	1.4k tok/s	56.8 tok/s	$12.32	$2.37	2.0 kW	Rent →
9	TPU v7 192GB	4	53 %	2.6k tok/s	99.3 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	94 %	~1.6k tok/s	87.1 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	77 %	1.4k tok/s	56.8 tok/s	$9.98	$1.93	2.8 kW	Rent →
12	B200 SXM 180GB	4	56 %	2.6k tok/s	104 tok/s	$16.36	$1.72	4.0 kW	Rent →
13	Ascend 970 announced	2	70 %	2.1k tok/s	102 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	4	53 %	2.6k tok/s	108 tok/s	—	—	2.8 kW	Specs →
15	Jalapeño announced	2	94 %	~2.1k tok/s	109 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	79 %	1.0k tok/s	42.6 tok/s	$7.60	$2.08	2.0 kW	Rent →
17	Instinct MI300X 192GB	4	53 %	1.7k tok/s	71.3 tok/s	$9.56	$1.52	3.0 kW	Rent →
18	BR100	8	80 %	1.2k tok/s	40.7 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	70 %	1.3k tok/s	68.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	72 %	1.5k tok/s	64.6 tok/s	$9.92	$1.82	2.8 kW	Rent →
21	H100 SXM5 80GB	8	64 %	2.2k tok/s	85.2 tok/s	$14.64	$1.87	5.6 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	82 %	2.1k tok/s	73.5 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	79 %	1.1k tok/s	44.1 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	72 %	~1.4k tok/s	64.6 tok/s	$14.46	$2.78	2.4 kW	Rent →
25	H100 NVL 94GB	8	54 %	~2.3k tok/s	99.2 tok/s	$25.52	$3.02	3.2 kW	Rent →
26	Ascend 910C	4	79 %	1.1k tok/s	43.1 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	8	64 %	1.4k tok/s	50.9 tok/s	$20.00	$4.09	2.8 kW	Rent →
28	H800 PCIe 80GB	8	64 %	1.4k tok/s	50.9 tok/s	—	—	2.8 kW	Specs →
29	Ascend 950DT	4	70 %	1.1k tok/s	53.8 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	8	53 %	1.1k tok/s	45.6 tok/s	$14.40	$3.48	4.8 kW	Rent →
31	Ascend 950PR	4	79 %	555 tok/s	21.5 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	8	64 %	~1.0k tok/s	40.7 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	8	53 %	1.0k tok/s	40.6 tok/s	$5.52	$1.46	4.8 kW	Rent →
34	Gaudi 3 128GB	4	79 %	1.4k tok/s	49.8 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	8	54 %	1.5k tok/s	70.3 tok/s	—	—	3.6 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	53 %	1.1k tok/s	45.6 tok/s	$4.00	$1.00	2.4 kW	Rent →
37	Cloud AI 100 Ultra	4	79 %	~216 tok/s	7.4 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	8	53 %	1.4k tok/s	62.3 tok/s	$8.80	$1.75	4.8 kW	Rent →
39	Atlas 350	4	91 %	483 tok/s	18.8 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB 2 nodes	16	54 %	1.4k tok/s	55.1 tok/s	—	—	4.8 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	82 %	1.6k tok/s	80.4 tok/s	$19.04	$3.38	9.2 kW	Rent →
42	Ascend 910B	8	80 %	1.0k tok/s	40.7 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	79 %	858 tok/s	44.1 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB 2 nodes	16	54 %	~1.1k tok/s	43.1 tok/s	$15.52	$3.75	4.8 kW	Rent →
45	Instinct MI250 128GB	4	79 %	839 tok/s	44.1 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB 2 nodes	16	54 %	1.1k tok/s	38.8 tok/s	$12.16	$3.21	5.6 kW	Rent →
47	A800 PCIe 80GB	8	64 %	1.1k tok/s	49.2 tok/s	—	—	2.4 kW	Specs →
48	A800 SXM4	8	64 %	1.1k tok/s	51.8 tok/s	—	—	3.2 kW	Specs →
49	A100 PCIe 80GB	8	64 %	1.1k tok/s	49.2 tok/s	$10.80	$2.80	2.4 kW	Rent →
50	A100 SXM4 80GB	8	64 %	1.1k tok/s	51.8 tok/s	$11.20	$2.80	3.2 kW	Rent →
51	A100 PCIe 40GB 2 nodes	16	65 %	1.6k tok/s	69.8 tok/s	—	—	4.0 kW	Specs →
52	A100 SXM4 40GB 2 nodes	16	65 %	1.6k tok/s	69.8 tok/s	$24.98	$4.27	6.4 kW	Rent →
53	RTX 5880 Ada 48GB 2 nodes	16	54 %	~1.1k tok/s	43.1 tok/s	—	—	4.6 kW	Specs →
54	TPU v4 32GB 2 nodes	16	82 %	1.3k tok/s	53.8 tok/s	—	—	—	Specs →
55	RTX 5000 Ada 32GB 2 nodes	16	82 %	~709 tok/s	25.8 tok/s	$13.28	$5.20	4.0 kW	Rent →
56	RTX PRO 5000 Blackwell 48GB 2 nodes	16	54 %	~1.4k tok/s	60.3 tok/s	$15.36	$3.08	4.8 kW	Rent →
57	RTX PRO 5000 Blackwell 72GB	8	71 %	~785 tok/s	34.2 tok/s	—	—	2.4 kW	Specs →
58	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	82 %	886 tok/s	35.9 tok/s	—	—	2.6 kW	Specs →
59	RTX PRO 4500 Blackwell 32GB 2 nodes	16	82 %	965 tok/s	40.2 tok/s	$11.52	$3.31	3.2 kW	Rent →
60	Radeon AI PRO R9700 32GB 2 nodes	16	82 %	734 tok/s	28.7 tok/s	—	—	4.8 kW	Specs →
61	Inferentia2 2 nodes	16	82 %	888 tok/s	36.8 tok/s	—	—	—	Specs →
62	Trainium 2 nodes	16	82 %	888 tok/s	36.8 tok/s	—	—	—	Specs →
63	Instinct MI100 32GB 2 nodes	16	82 %	881 tok/s	55.1 tok/s	—	—	4.8 kW	Specs →
64	Arc Pro B70 32GB 2 nodes	16	82 %	~699 tok/s	27.3 tok/s	—	—	3.7 kW	Specs →
65	L40 48GB 2 nodes	16	54 %	912 tok/s	38.8 tok/s	$13.12	$3.99	4.8 kW	Rent →
66	Instinct MI210 PCIe	8	80 %	792 tok/s	41.7 tok/s	—	—	2.4 kW	Specs →
67	RTX A6000 48GB 2 nodes	16	54 %	803 tok/s	34.5 tok/s	$8.00	$2.77	4.8 kW	Rent →
68	A40 48GB 2 nodes	16	54 %	740 tok/s	31.2 tok/s	$9.44	$3.54	4.8 kW	Rent →
69	H20 141GB HBM3e	4	72 %	567 tok/s	53.8 tok/s	—	—	1.6 kW	Specs →
70	H20 96GB	8	53 %	1.1k tok/s	102 tok/s	—	—	3.2 kW	Specs →
71	BI-V100 Tiangai 100 2 nodes	16	82 %	1.1k tok/s	53.8 tok/s	—	—	4.0 kW	Specs →
72	Atlas 300I Duo 48GB 2 nodes	16	54 %	480 tok/s	18.3 tok/s	—	—	2.4 kW	Specs →
73	Atlas 300I Duo 96GB	8	53 %	272 tok/s	10.4 tok/s	—	—	1.2 kW	Specs →
74	V100S PCIe 32GB 2 nodes	16	82 %	982 tok/s	50.9 tok/s	$14.08	$3.98	4.0 kW	Rent →
75	ATOM-Max	8	80 %	520 tok/s	26.0 tok/s	—	—	2.8 kW	Specs →
76	Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	54 %	811 tok/s	38.8 tok/s	—	—	4.7 kW	Specs →
77	Radeon PRO W7900 48GB 2 nodes	16	54 %	811 tok/s	38.8 tok/s	—	—	4.7 kW	Specs →
78	L20 48GB 2 nodes	16	54 %	803 tok/s	38.8 tok/s	—	—	4.4 kW	Specs →
79	Radeon AI PRO R9600D 32GB 2 nodes	16	82 %	618 tok/s	28.7 tok/s	—	—	2.4 kW	Specs →
80	Arc Pro B65 32GB 2 nodes	16	82 %	~596 tok/s	27.3 tok/s	—	—	3.2 kW	Specs →
81	MLU370-X8 2 nodes	16	54 %	595 tok/s	27.6 tok/s	—	—	4.0 kW	Specs →
82	Radeon PRO W7800 48GB 2 nodes	16	54 %	721 tok/s	38.8 tok/s	—	—	4.2 kW	Specs →
83	Radeon PRO W7800 32GB 2 nodes	16	82 %	559 tok/s	25.8 tok/s	—	—	4.2 kW	Specs →
84	Atlas 300V Pro 2 nodes	16	54 %	241 tok/s	9.2 tok/s	—	—	—	Specs →
85	Radeon PRO W6800 32GB 2 nodes	16	82 %	353 tok/s	23.0 tok/s	—	—	4.0 kW	Specs →
86	Radeon PRO V710 28GB 2 nodes	16	94 %	290 tok/s	20.1 tok/s	—	—	2.5 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 1.1k tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 1.7k tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 1.52x the tokens of AMD here.

NVIDIA 37 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 1.1k tok/s for $4.00/hr

Fewest GPUs: 2 x Rubin SXM, 3.5k tok/s

AMD 17 parts fit

Best value: 4 x Instinct MI300X 192GB, 1.7k tok/s for $9.56/hr

Fewest GPUs: 1 x Instinct MI455X, 2.1k tok/s

Huawei 8 parts fit

Fewest GPUs: 4 x Ascend 950DT, 1.1k tok/s

Intel 6 parts fit

Best value: 8 x Gaudi 2 96GB, 1.4k tok/s for $8.80/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 1.4k tok/s

Google 4 parts fit

Fewest GPUs: 4 x TPU v7 192GB, 2.6k tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 888 tok/s

Other 6 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 216 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	379k tok/s	$2,480	700 kW	$1.82
B	100 x GB200 NVL72 GPU 186GB	25 x TP4	0	69k tok/s	$1,050	—	$4.20

Size Qwen3.8-Flash-Next 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 Qwen3.8-Flash-Next need?

The weights take 360 GB at the native BF16 precision (360 GB at BF16, 180 GB at FP8, 101 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 Qwen3.8-Flash-Next?

2 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 Qwen3.8-Flash-Next?

8 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 8 H100s, and 1,000 tokens per second takes 8 H100s across 1 replicas.

How much does Qwen3.8-Flash-Next cost per million tokens on an H100?

About $1.87 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 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00.

What precision does Qwen3.8-Flash-Next 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. Experimental architecture: 125B of transformer weights with 6B active, plus a 51B n-gram embedding table and a 4B prediction layer, all counted in memory here. A sparse-attention indexer rides on the 12 full-attention layers and is not counted. Active parameters per the Qwen model card.