Moonshot AI Kimi K2

Kimi K2.5 hardware requirements

Needs at least 4 GPUs at INT4 / FP4 (GB300 NVL72 GPU 288GB). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00/hr.

Parameters

1.03T

Active per token

32.0B

mixture of experts

KV cache per token

69 KB

16-bit cache

Native precision

FP4

Rent Model card

Top options for Kimi K2.5

The strongest part each vendor offers for Kimi K2.5 at INT4 / FP4 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 GB300 NVL72 GPU 288GB 4 x Instinct MI355X OAM 8 x Ascend 950DT 8 x Gaudi 2 96GB 4 x TPU v7 192GB

Memory fills at: 4 x GB300 NVL72 GPU 288GB 230 users, 4 x Instinct MI355X OAM 230 users, 8 x Ascend 950DT 228 users, 8 x Gaudi 2 96GB 67 users, 4 x TPU v7 192GB 69 users. Roofline estimates at INT4 / FP4.

Bare minimum

Fewest GPUs

4 x GB300 NVL72 GPU 288GB

INT4 / FP4, 4,096 tokens, one stream

Tokens/s: 1.2k
Per hour: $49.68
Per million tokens: $11.78
Memory used: 52%

Rent on DataCrunch

Cheapest per hour

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 4,096 tokens, one stream, live price

Tokens/s: 492
Per hour: $4.00
Per million tokens: $2.26
Memory used: 79%

Rent on RunPod

Cheapest per million tokens

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 32,768 tokens, 32 streams

Tokens/s: 661
Per hour: $4.00
Per million tokens: $1.68
Memory used: 89%

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; 41 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	2	78 %	2.6k tok/s	92.2 tok/s	—	—	—	Specs →
2	Rubin SXM	4	59 %	4.6k tok/s	165 tok/s	—	—	—	Specs →
3	TPU 8t announced	4	78 %	~1.4k tok/s	49.0 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	59 %	~1.7k tok/s	64.5 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	59 %	1.6k tok/s	60.0 tok/s	$23.96	$4.08	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	59 %	1.7k tok/s	60.0 tok/s	$49.68	$8.04	—	Rent →
7	GB200 NVL72 GPU 186GB	4	91 %	1.6k tok/s	60.0 tok/s	$42.00	$7.16	—	Rent →
8	Instinct MI350X OAM	4	59 %	1.6k tok/s	60.0 tok/s	$24.64	$4.26	4.0 kW	Rent →
9	TPU v7 192GB	4	88 %	~1.0k tok/s	55.4 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	2	78 %	~1.9k tok/s	92.2 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	64 %	1.7k tok/s	60.0 tok/s	$19.96	$3.27	5.6 kW	Rent →
12	B200 SXM 180GB	4	94 %	1.6k tok/s	57.8 tok/s	$16.36	$2.93	4.0 kW	Rent →
13	Ascend 970 announced	4	59 %	2.5k tok/s	108 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	4	88 %	1.5k tok/s	60.0 tok/s	—	—	2.8 kW	Specs →
15	Jalapeño announced	4	78 %	~2.9k tok/s	116 tok/s	—	—	2.8 kW	Specs →
16	Instinct MI325X OAM	4	66 %	~710 tok/s	45.0 tok/s	$15.20	$5.95	4.0 kW	Rent →
17	Instinct MI300X 192GB	4	88 %	~667 tok/s	39.8 tok/s	$9.56	$3.98	3.0 kW	Rent →
18	BR100 2 nodes	16	67 %	~948 tok/s	40.0 tok/s	—	—	8.8 kW	Specs →
19	Ascend 960 announced	4	59 %	1.5k tok/s	72.0 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	8	60 %	~1.0k tok/s	68.0 tok/s	$19.84	$5.28	5.6 kW	Rent →
21	H100 SXM5 80GB 2 nodes	16	53 %	~1.5k tok/s	83.8 tok/s	$29.28	$5.33	11.2 kW	Rent →
22	Data Center GPU Max 1550 128GB	8	66 %	~792 tok/s	46.4 tok/s	—	—	4.8 kW	Specs →
23	H200 NVL 141GB	8	60 %	~952 tok/s	68.0 tok/s	$28.91	$8.44	4.8 kW	Rent →
24	H100 NVL 94GB	8	91 %	~864 tok/s	55.3 tok/s	$25.52	$8.20	3.2 kW	Rent →
25	Ascend 910C	8	66 %	~765 tok/s	45.3 tok/s	—	—	4.8 kW	Specs →
26	H100 PCIe 80GB 2 nodes	16	53 %	~1.0k tok/s	50.0 tok/s	$40.00	$11.06	5.6 kW	Rent →
27	H800 PCIe 80GB 2 nodes	16	53 %	~1.0k tok/s	50.0 tok/s	—	—	5.6 kW	Specs →
28	Ascend 950DT	8	59 %	1.3k tok/s	56.7 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition	8	89 %	678 tok/s	25.4 tok/s	$14.40	$5.90	4.8 kW	Rent →
30	Ascend 950PR	8	66 %	615 tok/s	22.7 tok/s	—	—	7.2 kW	Specs →
31	MTT S5000 2 nodes	16	53 %	~746 tok/s	40.0 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition	8	89 %	614 tok/s	22.6 tok/s	$5.52	$2.50	4.8 kW	Rent →
33	Gaudi 3 128GB	8	66 %	~1.1k tok/s	52.4 tok/s	—	—	7.2 kW	Specs →
34	TPU v5p 95GB	8	90 %	~534 tok/s	39.2 tok/s	—	—	3.6 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	89 %	661 tok/s	25.4 tok/s	$4.00	$1.68	2.4 kW	Rent →
36	Cloud AI 100 Ultra	8	66 %	~194 tok/s	7.8 tok/s	—	—	1.2 kW	Specs →
37	Gaudi 2 96GB	8	89 %	~489 tok/s	34.7 tok/s	$8.80	$5.00	4.8 kW	Rent →
38	Atlas 350	8	76 %	529 tok/s	19.8 tok/s	—	—	4.8 kW	Specs →
39	Data Center GPU Max 1100 48GB 2 nodes	16	90 %	~594 tok/s	30.7 tok/s	—	—	4.8 kW	Specs →
40	Ascend 910B 2 nodes	16	67 %	~675 tok/s	40.0 tok/s	—	—	6.4 kW	Specs →
41	MI250X 128GB	8	66 %	~509 tok/s	46.4 tok/s	—	—	4.5 kW	Specs →
42	RTX 6000 Ada 48GB 2 nodes	16	90 %	~483 tok/s	24.0 tok/s	$15.52	$8.93	4.8 kW	Rent →
43	Instinct MI250 128GB	8	66 %	~491 tok/s	46.4 tok/s	—	—	4.0 kW	Specs →
44	L40S 48GB 2 nodes	16	90 %	~450 tok/s	21.6 tok/s	$12.16	$7.50	5.6 kW	Rent →
45	A800 PCIe 80GB 2 nodes	16	53 %	~648 tok/s	48.4 tok/s	—	—	4.8 kW	Specs →
46	A800 SXM4 2 nodes	16	53 %	~662 tok/s	51.0 tok/s	—	—	6.4 kW	Specs →
47	A100 PCIe 80GB 2 nodes	16	53 %	~648 tok/s	48.4 tok/s	$21.60	$9.26	4.8 kW	Rent →
48	A100 SXM4 80GB 2 nodes	16	53 %	~662 tok/s	51.0 tok/s	$22.40	$9.40	6.4 kW	Rent →
49	RTX 5880 Ada 48GB 2 nodes	16	90 %	~432 tok/s	24.0 tok/s	—	—	4.6 kW	Specs →
50	RTX PRO 5000 Blackwell 48GB 2 nodes	16	90 %	~832 tok/s	33.6 tok/s	$15.36	$5.13	4.8 kW	Rent →
51	RTX PRO 5000 Blackwell 72GB 2 nodes	16	59 %	~832 tok/s	33.6 tok/s	—	—	4.8 kW	Specs →
52	L40 48GB 2 nodes	16	90 %	~334 tok/s	21.6 tok/s	$13.12	$10.91	4.8 kW	Rent →
53	Instinct MI210 PCIe 2 nodes	16	67 %	~433 tok/s	41.0 tok/s	—	—	4.8 kW	Specs →
54	RTX A6000 48GB 2 nodes	16	90 %	~291 tok/s	19.2 tok/s	$8.00	$7.64	4.8 kW	Rent →
55	A40 48GB 2 nodes	16	90 %	~273 tok/s	17.4 tok/s	$9.44	$9.61	4.8 kW	Rent →
56	H20 141GB HBM3e	8	60 %	~257 tok/s	56.7 tok/s	—	—	3.2 kW	Specs →
57	H20 96GB	8	89 %	~257 tok/s	56.7 tok/s	—	—	3.2 kW	Specs →
58	Atlas 300I Duo 48GB 2 nodes	16	90 %	~198 tok/s	10.2 tok/s	—	—	2.4 kW	Specs →
59	Atlas 300I Duo 96GB	8	89 %	~112 tok/s	5.8 tok/s	—	—	1.2 kW	Specs →
60	ATOM-Max 2 nodes	16	67 %	~293 tok/s	25.6 tok/s	—	—	5.6 kW	Specs →
61	Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	90 %	~269 tok/s	21.6 tok/s	—	—	4.7 kW	Specs →
62	Radeon PRO W7900 48GB 2 nodes	16	90 %	~269 tok/s	21.6 tok/s	—	—	4.7 kW	Specs →
63	L20 48GB 2 nodes	16	90 %	~264 tok/s	21.6 tok/s	—	—	4.4 kW	Specs →
64	MLU370-X8 2 nodes	16	90 %	~202 tok/s	15.4 tok/s	—	—	4.0 kW	Specs →
65	Radeon PRO W7800 48GB 2 nodes	16	90 %	~220 tok/s	21.6 tok/s	—	—	4.2 kW	Specs →
66	Atlas 300V Pro 2 nodes	16	90 %	~99.0 tok/s	5.1 tok/s	—	—	—	Specs →

By vendor

At INT4 / FP4 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 661 tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 667 tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 2.37x the tokens of AMD here.

NVIDIA 30 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 661 tok/s for $4.00/hr

Fewest GPUs: 4 x Rubin SXM, 4.6k tok/s

AMD 11 parts fit

Best value: 4 x Instinct MI300X 192GB, 667 tok/s for $9.56/hr

Fewest GPUs: 2 x Instinct MI455X, 2.6k tok/s

Huawei 8 parts fit

Fewest GPUs: 8 x Ascend 950DT, 1.3k tok/s

Intel 4 parts fit

Best value: 8 x Gaudi 2 96GB, 489 tok/s for $8.80/hr

Fewest GPUs: 8 x Gaudi 3 128GB, 1.1k tok/s

Google 2 parts fit

Fewest GPUs: 4 x TPU v7 192GB, 1.0k tok/s

Other 5 parts fit

Fewest GPUs: 8 x Cloud AI 100 Ultra, 194 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	130k tok/s	$2,480	700 kW	$5.28
B	100 x GB200 NVL72 GPU 186GB	25 x TP4	0	41k tok/s	$1,050	—	$7.16

Size Kimi K2.5 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 Kimi K2.5 need?

The weights take 578 GB at the native INT4 / FP4 precision (2,054 GB at BF16, 1,027 GB at FP8, 578 GB at INT4). Each concurrent stream adds 69 KB of KV cache per token: 0.3 GB at 4,096 tokens and 2.3 GB at 32,768 tokens.

What is the bare minimum to run Kimi K2.5?

4 x GB300 NVL72 GPU 288GB at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00 per hour.

How many H100s do you need to run Kimi K2.5?

8 H100s at INT4 / FP4 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 Kimi K2.5 cost per million tokens on an H100?

About $5.33 per million output tokens at INT4 / FP4, 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 $1.68.

What precision does Kimi K2.5 ship in?

The published checkpoint is FP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.

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. Kimi K2 geometry, released as INT4 quantisation-aware weights (group size 32), so the INT4 column is the native one. Multimodal. Active parameters per the Moonshot model card.