MiniMax MiniMax M3

MiniMax M3 hardware requirements

Needs at least 4 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 32 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $16.00/hr.

Parameters

427B

Active per token

23.0B

mixture of experts

KV cache per token

120 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for MiniMax M3

The strongest part each vendor offers for MiniMax M3 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.

8 x GB200 NVL72 GPU 186GB 4 x Instinct MI355X OAM 8 x Ascend 950DT 16 x Gaudi 2 96GB 8 x TPU v7 192GB

Memory fills at: 8 x GB200 NVL72 GPU 186GB 142 users, 4 x Instinct MI355X OAM 63 users, 8 x Ascend 950DT 62 users, 16 x Gaudi 2 96GB 152 users, 8 x TPU v7 192GB 154 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

4 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 458
Per hour: $23.96
Per million tokens: $14.53
Memory used: 77%

Rent on Daytona

Cheapest per hour

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 341
Per hour: $8.00
Per million tokens: $6.52
Memory used: 58%

Rent on RunPod

Cheapest per million tokens

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 562
Per hour: $8.00
Per million tokens: $3.95
Memory used: 67%

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; 28 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	4	59 %	2.1k tok/s	92.4 tok/s	—	—	—	Specs →
2	Rubin SXM	4	89 %	1.9k tok/s	87.2 tok/s	—	—	—	Specs →
3	TPU 8t announced	8	59 %	~1.3k tok/s	48.9 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	89 %	~846 tok/s	34.1 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	89 %	799 tok/s	31.7 tok/s	$23.96	$8.33	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	89 %	798 tok/s	31.7 tok/s	$49.68	$17.30	—	Rent →
7	GB200 NVL72 GPU 186GB	8	69 %	1.5k tok/s	59.9 tok/s	$84.00	$15.48	—	Rent →
8	Instinct MI350X OAM	4	89 %	784 tok/s	31.7 tok/s	$24.64	$8.73	4.0 kW	Rent →
9	TPU v7 192GB	8	67 %	1.4k tok/s	55.3 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	4	59 %	~1.6k tok/s	92.4 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	97 %	779 tok/s	31.7 tok/s	$19.96	$7.12	5.6 kW	Rent →
12	B200 SXM 180GB	8	71 %	1.4k tok/s	57.7 tok/s	$32.72	$6.36	8.0 kW	Rent →
13	Ascend 970 announced	4	89 %	1.1k tok/s	57.1 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	8	67 %	1.4k tok/s	59.9 tok/s	—	—	5.6 kW	Specs →
15	Jalapeño announced	8	59 %	~2.1k tok/s	115 tok/s	—	—	5.6 kW	Specs →
16	Instinct MI325X OAM	4	100 %	547 tok/s	23.8 tok/s	$15.20	$7.71	4.0 kW	Rent →
17	Instinct MI300X 192GB	8	67 %	944 tok/s	39.7 tok/s	$19.12	$5.62	6.0 kW	Rent →
18	Ascend 960 announced	4	89 %	670 tok/s	38.1 tok/s	—	—	—	Specs →
19	H200 SXM 141GB	8	91 %	814 tok/s	36.0 tok/s	$19.84	$6.77	5.6 kW	Rent →
20	H100 SXM5 80GB 2 nodes	16	81 %	1.1k tok/s	44.3 tok/s	$29.28	$7.39	11.2 kW	Rent →
21	Data Center GPU Max 1550 128GB 2 nodes	16	50 %	1.0k tok/s	43.3 tok/s	—	—	9.6 kW	Specs →
22	H200 NVL 141GB	8	91 %	~773 tok/s	36.0 tok/s	$28.91	$10.39	4.8 kW	Rent →
23	H100 NVL 94GB 2 nodes	16	68 %	~1.2k tok/s	51.5 tok/s	$51.04	$12.01	6.4 kW	Rent →
24	Ascend 910C 2 nodes	16	50 %	1.0k tok/s	42.3 tok/s	—	—	9.6 kW	Specs →
25	H100 PCIe 80GB 2 nodes	16	81 %	691 tok/s	26.4 tok/s	$40.00	$16.09	5.6 kW	Rent →
26	H800 PCIe 80GB 2 nodes	16	81 %	691 tok/s	26.4 tok/s	—	—	5.6 kW	Specs →
27	Ascend 950DT	8	89 %	591 tok/s	30.0 tok/s	—	—	—	Specs →
28	RTX PRO 6000 Blackwell Workstation Edition 2 nodes	16	67 %	582 tok/s	23.7 tok/s	$28.80	$13.75	9.6 kW	Rent →
29	Ascend 950PR 2 nodes	16	50 %	532 tok/s	21.1 tok/s	—	—	14.4 kW	Specs →
30	MTT S5000 2 nodes	16	81 %	~532 tok/s	21.1 tok/s	—	—	—	Specs →
31	RTX PRO 6000 Blackwell Server Edition 2 nodes	16	67 %	531 tok/s	21.1 tok/s	$11.04	$5.77	9.6 kW	Rent →
32	Gaudi 3 128GB 2 nodes	16	50 %	1.3k tok/s	48.9 tok/s	—	—	14.4 kW	Specs →
33	TPU v5p 95GB 2 nodes	16	68 %	773 tok/s	36.5 tok/s	—	—	7.2 kW	Specs →
34	RTX PRO 6000 Blackwell Max-Q Workstation Edition 2 nodes	16	67 %	562 tok/s	23.7 tok/s	$8.00	$3.95	4.8 kW	Rent →
35	Cloud AI 100 Ultra 2 nodes	16	50 %	~209 tok/s	7.2 tok/s	—	—	2.4 kW	Specs →
36	Gaudi 2 96GB 2 nodes	16	67 %	699 tok/s	32.4 tok/s	$17.60	$7.00	9.6 kW	Rent →
37	Atlas 350 2 nodes	16	57 %	462 tok/s	18.5 tok/s	—	—	9.6 kW	Specs →
38	MI250X 128GB 2 nodes	16	50 %	802 tok/s	43.3 tok/s	—	—	9.0 kW	Specs →
39	Instinct MI250 128GB 2 nodes	16	50 %	783 tok/s	43.3 tok/s	—	—	8.0 kW	Specs →
40	A800 PCIe 80GB 2 nodes	16	81 %	536 tok/s	25.6 tok/s	—	—	4.8 kW	Specs →
41	A800 SXM4 2 nodes	16	81 %	554 tok/s	27.0 tok/s	—	—	6.4 kW	Specs →
42	A100 PCIe 80GB 2 nodes	16	81 %	536 tok/s	25.6 tok/s	$21.60	$11.20	4.8 kW	Rent →
43	A100 SXM4 80GB 2 nodes	16	81 %	554 tok/s	27.0 tok/s	$22.40	$11.23	6.4 kW	Rent →
44	RTX PRO 5000 Blackwell 72GB 2 nodes	16	90 %	~394 tok/s	17.8 tok/s	—	—	4.8 kW	Specs →
45	H20 141GB HBM3e	8	91 %	290 tok/s	30.0 tok/s	—	—	3.2 kW	Specs →
46	H20 96GB 2 nodes	16	67 %	513 tok/s	52.9 tok/s	—	—	6.4 kW	Specs →
47	Atlas 300I Duo 96GB 2 nodes	16	67 %	138 tok/s	5.4 tok/s	—	—	2.4 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 562 tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 944 tokens/s for $19.12/hr. Per rental dollar NVIDIA delivers 1.42x the tokens of AMD here.

NVIDIA 22 parts fit

Best value: 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 562 tok/s for $8.00/hr

Fewest GPUs: 4 x Rubin SXM, 1.9k tok/s

AMD 7 parts fit

Best value: 8 x Instinct MI300X 192GB, 944 tok/s for $19.12/hr

Fewest GPUs: 4 x Instinct MI455X, 2.1k tok/s

Huawei 5 parts fit

Fewest GPUs: 8 x Ascend 950DT, 591 tok/s

Intel 3 parts fit

Best value: 16 x Gaudi 2 96GB, 699 tok/s for $17.60/hr

Fewest GPUs: 16 x Gaudi 3 128GB, 1.3k tok/s

Google 2 parts fit

Fewest GPUs: 8 x TPU v7 192GB, 1.4k tok/s

Other 2 parts fit

Fewest GPUs: 16 x MTT S5000, 532 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	102k tok/s	$2,480	700 kW	$6.77
B	104 x GB200 NVL72 GPU 186GB	13 x TP8	0	20k tok/s	$1,092	—	$15.48

Size MiniMax M3 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 MiniMax M3 need?

The weights take 854 GB at the native BF16 precision (854 GB at BF16, 427 GB at FP8, 240 GB at INT4). Each concurrent stream adds 120 KB of KV cache per token: 0.5 GB at 4,096 tokens and 4.0 GB at 32,768 tokens.

What is the bare minimum to run MiniMax M3?

4 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00 per hour.

How many H100s do you need to run MiniMax M3?

16 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 16 H100s, and 1,000 tokens per second takes 16 H100s across 1 replicas.

How much does MiniMax M3 cost per million tokens on an H100?

About $7.39 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 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $3.95.

What precision does MiniMax M3 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. Multimodal, with block-sparse attention that reads a top-k subset of the cache; the cache itself still holds every token (120 KB per token) and the small index cache is not counted. Active parameters are the cards approximate figure (about 23B of about 428B).