Google Gemma 4

Gemma 4 26B-A4B 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 3 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.50/hr.

Parameters

25.8B

Active per token

4.0B

mixture of experts

KV cache per token

110 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Gemma 4 26B-A4B

The strongest part each vendor offers for Gemma 4 26B-A4B 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 910C 1 x Gaudi 2 96GB 1 x TPU v7 192GB 2 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 512 users, 1 x Instinct MI355X OAM 512 users, 1 x Ascend 910C 161 users, 1 x Gaudi 2 96GB 91 users, 1 x TPU v7 192GB 301 users, 2 x Inferentia2 18 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 727
Per hour: $5.99
Per million tokens: $2.29
Memory used: 19%

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: 162
Per hour: $0.50
Per million tokens: $0.86
Memory used: 56%

Rent on RunPod

Cheapest per million tokens

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 426
Per hour: $0.50
Per million tokens: $0.33
Memory used: 72%

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	16 %	5.3k tok/s	292 tok/s	—	—	—	Specs →
2	Rubin SXM	1	24 %	4.6k tok/s	276 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	32 %	~1.9k tok/s	81.9 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	24 %	~2.2k tok/s	108 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	24 %	2.1k tok/s	100 tok/s	$5.99	$0.80	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	24 %	2.1k tok/s	100 tok/s	$12.42	$1.65	—	Rent →
7	GB200 NVL72 GPU 186GB	1	37 %	2.1k tok/s	100 tok/s	$10.50	$1.40	—	Rent →
8	Instinct MI350X OAM	1	24 %	2.0k tok/s	100 tok/s	$6.16	$0.84	1.0 kW	Rent →
9	TPU v7 192GB	1	36 %	1.9k tok/s	92.5 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	16 %	~3.5k tok/s	292 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	26 %	2.0k tok/s	100 tok/s	$4.99	$0.69	1.4 kW	Rent →
12	B200 SXM 180GB	1	38 %	2.0k tok/s	96.6 tok/s	$4.09	$0.58	1.0 kW	Rent →
13	Ascend 970 announced	1	24 %	2.6k tok/s	181 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	36 %	1.8k tok/s	100 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	32 %	~2.4k tok/s	193 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	27 %	1.4k tok/s	75.2 tok/s	$3.80	$0.78	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	36 %	1.3k tok/s	66.5 tok/s	$2.39	$0.53	0.8 kW	Rent →
18	BR100	2	54 %	966 tok/s	38.1 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	24 %	1.5k tok/s	120 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	49 %	1.1k tok/s	60.2 tok/s	$2.48	$0.65	0.7 kW	Rent →
21	H100 SXM5 80GB	1	86 %	856 tok/s	42.0 tok/s	$1.83	$0.59	0.7 kW	Rent →
22	TPU v6e 32GB	4	55 %	1.8k tok/s	73.9 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	54 %	793 tok/s	41.1 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	49 %	~979 tok/s	60.2 tok/s	$3.61	$1.03	0.6 kW	Rent →
25	H100 NVL 94GB	1	73 %	~876 tok/s	48.9 tok/s	$3.19	$1.01	0.4 kW	Rent →
26	Ascend 910C	1	54 %	767 tok/s	40.1 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	1	86 %	555 tok/s	25.1 tok/s	$2.50	$1.25	0.3 kW	Rent →
28	H800 PCIe 80GB	1	86 %	555 tok/s	25.1 tok/s	—	—	0.3 kW	Specs →
29	Ascend 950DT	1	48 %	719 tok/s	50.2 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	72 %	450 tok/s	22.5 tok/s	$1.80	$1.11	0.6 kW	Rent →
31	Ascend 950PR	1	54 %	417 tok/s	20.1 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	1	86 %	~417 tok/s	20.1 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	72 %	417 tok/s	20.0 tok/s	$0.69	$0.46	0.6 kW	Rent →
34	Gaudi 3 128GB	1	54 %	1.1k tok/s	46.4 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	72 %	551 tok/s	34.7 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	72 %	426 tok/s	22.5 tok/s	$0.50	$0.33	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	54 %	~181 tok/s	6.9 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	72 %	503 tok/s	30.7 tok/s	$1.10	$0.61	0.6 kW	Rent →
39	Atlas 350	1	61 %	361 tok/s	17.6 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	2	73 %	627 tok/s	29.3 tok/s	—	—	0.6 kW	Specs →
41	GeForce RTX 5090 32GB	4	55 %	1.1k tok/s	80.9 tok/s	$4.76	$1.25	2.3 kW	Rent →
42	Ascend 910B	2	54 %	729 tok/s	38.1 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	54 %	538 tok/s	41.1 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	2	73 %	~507 tok/s	22.9 tok/s	$1.94	$1.06	0.6 kW	Rent →
45	Instinct MI250 128GB	1	54 %	521 tok/s	41.1 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	2	73 %	470 tok/s	20.6 tok/s	$1.52	$0.90	0.7 kW	Rent →
47	A800 PCIe 80GB	1	86 %	380 tok/s	24.3 tok/s	—	—	0.3 kW	Specs →
48	A800 SXM4	1	86 %	389 tok/s	25.6 tok/s	—	—	0.4 kW	Specs →
49	A100 PCIe 80GB	1	86 %	380 tok/s	24.3 tok/s	$1.35	$0.99	0.3 kW	Rent →
50	A100 SXM4 80GB	1	86 %	389 tok/s	25.6 tok/s	$1.40	$1.00	0.4 kW	Rent →
51	A100 PCIe 40GB	2	87 %	644 tok/s	37.1 tok/s	—	—	0.5 kW	Specs →
52	A100 SXM4 40GB	2	87 %	644 tok/s	37.1 tok/s	$3.12	$1.35	0.8 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	4	74 %	~1.1k tok/s	60.7 tok/s	—	—	2.3 kW	Specs →
54	RTX 5880 Ada 48GB	2	73 %	~462 tok/s	22.9 tok/s	—	—	0.6 kW	Specs →
55	TPU v4 32GB	4	55 %	999 tok/s	54.2 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	55 %	~607 tok/s	26.0 tok/s	$3.32	$1.52	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	2	73 %	~546 tok/s	32.0 tok/s	$1.92	$0.98	0.6 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	1	96 %	~287 tok/s	16.9 tok/s	—	—	0.3 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	55 %	694 tok/s	36.1 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	55 %	742 tok/s	40.5 tok/s	$2.88	$1.08	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	55 %	591 tok/s	28.9 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	55 %	686 tok/s	37.0 tok/s	—	—	—	Specs →
63	Trainium	4	55 %	686 tok/s	37.0 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	55 %	547 tok/s	55.5 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	55 %	~564 tok/s	27.4 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	2	73 %	365 tok/s	20.6 tok/s	$1.64	$1.25	0.6 kW	Rent →
67	Instinct MI210 PCIe	2	54 %	495 tok/s	39.0 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	4	74 %	718 tok/s	45.5 tok/s	$3.56	$1.38	1.8 kW	Rent →
69	A30 24GB	4	74 %	690 tok/s	42.1 tok/s	—	—	0.7 kW	Specs →
70	RTX 4500 Ada 24GB	4	74 %	~428 tok/s	19.5 tok/s	—	—	0.8 kW	Specs →
71	RTX A6000 48GB	2	73 %	319 tok/s	18.3 tok/s	$1.00	$0.87	0.6 kW	Rent →
72	A40 48GB	2	73 %	298 tok/s	16.6 tok/s	$1.18	$1.10	0.6 kW	Rent →
73	H20 141GB HBM3e	1	49 %	289 tok/s	50.2 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	72 %	289 tok/s	50.2 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	4	74 %	~549 tok/s	30.3 tok/s	$2.28	$1.15	0.6 kW	Rent →
76	BI-V100 Tiangai 100	4	55 %	730 tok/s	54.2 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	2	73 %	209 tok/s	9.7 tok/s	—	—	0.3 kW	Specs →
78	Atlas 300I Duo 96GB	1	72 %	110 tok/s	5.1 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	4	74 %	~569 tok/s	34.7 tok/s	—	—	0.9 kW	Specs →
80	Wormhole n300s	4	74 %	478 tok/s	26.0 tok/s	—	—	1.2 kW	Specs →
81	V100S PCIe 32GB	4	55 %	663 tok/s	51.2 tok/s	$3.52	$1.47	1.0 kW	Rent →
82	ATOM-Max	2	54 %	332 tok/s	24.4 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	4	74 %	479 tok/s	27.1 tok/s	$4.41	$2.55	0.6 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	2	73 %	302 tok/s	20.6 tok/s	—	—	0.6 kW	Specs →
85	Radeon PRO W7900 48GB	2	73 %	302 tok/s	20.6 tok/s	—	—	0.6 kW	Specs →
86	Radeon RX 7900 XTX 24GB	4	74 %	598 tok/s	43.3 tok/s	—	—	1.4 kW	Specs →
87	L4 24GB	4	74 %	306 tok/s	13.5 tok/s	$2.36	$2.14	0.3 kW	Rent →
88	L20 48GB	2	73 %	297 tok/s	20.6 tok/s	—	—	0.6 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	55 %	442 tok/s	28.9 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	55 %	~430 tok/s	27.4 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	4	74 %	~370 tok/s	20.6 tok/s	—	—	0.8 kW	Specs →
92	L2 24GB	4	74 %	284 tok/s	13.5 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	4	74 %	~355 tok/s	19.5 tok/s	—	—	0.3 kW	Specs →
94	MLU370-X8	2	73 %	225 tok/s	14.6 tok/s	—	—	0.5 kW	Specs →
95	Radeon PRO W7800 48GB	2	73 %	252 tok/s	20.6 tok/s	—	—	0.5 kW	Specs →
96	Radeon PRO W7800 32GB	4	55 %	401 tok/s	26.0 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	4	74 %	420 tok/s	42.3 tok/s	$2.00	$1.32	1.4 kW	Rent →
98	A10G 24GB	4	74 %	355 tok/s	27.1 tok/s	$4.02	$3.15	1.2 kW	Rent →
99	Atlas 300V Pro	2	73 %	105 tok/s	4.9 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	4	74 %	198 tok/s	9.2 tok/s	—	—	—	Specs →
101	Atlas 300V	4	74 %	175 tok/s	9.2 tok/s	—	—	0.3 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	4	74 %	278 tok/s	45.5 tok/s	$1.84	$1.84	1.8 kW	Rent →
103	Radeon PRO W6800 32GB	4	55 %	216 tok/s	23.1 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	4	74 %	196 tok/s	34.7 tok/s	$1.08	$1.53	0.9 kW	Rent →
105	Radeon PRO V710 28GB	4	63 %	173 tok/s	20.2 tok/s	—	—	0.6 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 426 tokens/s for $0.50/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 1.61x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 426 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 4.6k 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 910C, 767 tok/s

Intel 7 parts fit

Best value: 1 x Gaudi 2 96GB, 503 tok/s for $1.10/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 1.1k 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, 686 tok/s

Other 7 parts fit

Fewest GPUs: 1 x MTT S5000, 417 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	1.06M tok/s	$2,480	700 kW	$0.65
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	209k tok/s	$1,050	—	$1.40

Size Gemma 4 26B-A4B 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 Gemma 4 26B-A4B need?

The weights take 51.6 GB at the native BF16 precision (51.6 GB at BF16, 25.8 GB at FP8, 14.5 GB at INT4). Each concurrent stream adds 110 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.4 GB at 32,768 tokens.

What is the bare minimum to run Gemma 4 26B-A4B?

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 Gemma 4 26B-A4B?

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 one H100, and 1,000 tokens per second takes 2 H100s across 2 replicas.

How much does Gemma 4 26B-A4B cost per million tokens on an H100?

About $0.59 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 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.33.

What precision does Gemma 4 26B-A4B 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. Active parameters per the model name (A4B); the expert geometry computes 4.4B. Five global layers, twenty-five sliding-window layers, K and V shared.