Ai2 OLMo 3

OLMo 3.1 32B Instruct hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.64/hr or 14 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $7.00/hr.

Parameters

32.2B

Active per token

All

dense model

KV cache per token

256 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for OLMo 3.1 32B Instruct

The strongest part each vendor offers for OLMo 3.1 32B Instruct 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 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 72 users, 1 x Instinct MI355X OAM 72 users, 1 x Ascend 910C 19 users, 1 x Gaudi 2 96GB 9 users, 1 x TPU v7 192GB 40 users, 4 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: 90.3
Per hour: $5.99
Per million tokens: $18.42
Memory used: 24%

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: 20.2
Per hour: $0.50
Per million tokens: $6.89
Memory used: 71%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 167
Per hour: $1.00
Per million tokens: $1.67
Memory used: 87%

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	38 %	1.0k tok/s	110 tok/s	—	—	—	Specs →
2	Rubin SXM	1	57 %	872 tok/s	104 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	77 %	~479 tok/s	30.8 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	57 %	~474 tok/s	40.6 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	57 %	460 tok/s	37.7 tok/s	$5.99	$3.62	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	57 %	458 tok/s	37.7 tok/s	$12.42	$7.53	—	Rent →
7	GB200 NVL72 GPU 186GB	1	89 %	458 tok/s	37.7 tok/s	$10.50	$6.36	—	Rent →
8	Instinct MI350X OAM	1	57 %	436 tok/s	37.7 tok/s	$6.16	$3.92	1.0 kW	Rent →
9	TPU v7 192GB	1	86 %	423 tok/s	34.8 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	38 %	~570 tok/s	110 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	63 %	429 tok/s	37.7 tok/s	$4.99	$3.23	1.4 kW	Rent →
12	B200 SXM 180GB	1	92 %	423 tok/s	36.3 tok/s	$4.09	$2.69	1.0 kW	Rent →
13	Ascend 970 announced	1	57 %	465 tok/s	67.9 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	86 %	362 tok/s	37.7 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	77 %	~413 tok/s	72.7 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	65 %	271 tok/s	28.3 tok/s	$3.80	$3.90	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	86 %	261 tok/s	25.0 tok/s	$2.39	$2.55	0.8 kW	Rent →
18	BR100	4	65 %	484 tok/s	27.2 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	57 %	245 tok/s	45.3 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	59 %	396 tok/s	43.0 tok/s	$4.96	$3.48	1.4 kW	Rent →
21	H100 SXM5 80GB	4	52 %	667 tok/s	56.9 tok/s	$7.32	$3.05	2.8 kW	Rent →
22	TPU v6e 32GB	8	67 %	880 tok/s	52.6 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	65 %	314 tok/s	29.4 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	2	59 %	~350 tok/s	43.0 tok/s	$7.23	$5.74	1.2 kW	Rent →
25	H100 NVL 94GB	2	89 %	~330 tok/s	35.0 tok/s	$6.38	$5.36	0.8 kW	Rent →
26	Ascend 910C	2	65 %	302 tok/s	28.7 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	4	52 %	462 tok/s	34.0 tok/s	$10.00	$6.01	1.4 kW	Rent →
28	H800 PCIe 80GB	4	52 %	462 tok/s	34.0 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	58 %	242 tok/s	35.9 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	87 %	182 tok/s	16.1 tok/s	$3.60	$5.48	1.2 kW	Rent →
31	Ascend 950PR	2	65 %	174 tok/s	14.3 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	4	52 %	~330 tok/s	27.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	87 %	174 tok/s	14.3 tok/s	$1.38	$2.20	1.2 kW	Rent →
34	Gaudi 3 128GB	2	65 %	499 tok/s	33.2 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	88 %	194 tok/s	24.8 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	87 %	167 tok/s	16.1 tok/s	$1.00	$1.67	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	65 %	~95.6 tok/s	4.9 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	87 %	180 tok/s	22.0 tok/s	$2.20	$3.39	1.2 kW	Rent →
39	Atlas 350	2	74 %	150 tok/s	12.6 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	88 %	268 tok/s	20.9 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	67 %	343 tok/s	57.5 tok/s	$9.52	$7.72	4.6 kW	Rent →
42	Ascend 910B	4	65 %	286 tok/s	27.2 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	65 %	175 tok/s	29.4 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	88 %	~222 tok/s	16.3 tok/s	$3.88	$4.85	1.2 kW	Rent →
45	Instinct MI250 128GB	2	65 %	167 tok/s	29.4 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	88 %	212 tok/s	14.7 tok/s	$3.04	$3.99	1.4 kW	Rent →
47	A800 PCIe 80GB	4	52 %	252 tok/s	32.9 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	52 %	255 tok/s	34.6 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	52 %	252 tok/s	32.9 tok/s	$5.40	$5.95	1.2 kW	Rent →
50	A100 SXM4 80GB	4	52 %	255 tok/s	34.6 tok/s	$5.60	$6.09	1.6 kW	Rent →
51	A100 PCIe 40GB	8	53 %	450 tok/s	49.9 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	53 %	450 tok/s	49.9 tok/s	$12.49	$7.71	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	90 %	~416 tok/s	43.1 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	88 %	~188 tok/s	16.3 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	67 %	381 tok/s	38.5 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	67 %	~278 tok/s	18.5 tok/s	$6.64	$6.63	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	88 %	~199 tok/s	22.8 tok/s	$3.84	$5.35	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	58 %	~199 tok/s	22.8 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	67 %	273 tok/s	25.7 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	67 %	283 tok/s	28.7 tok/s	$5.76	$5.66	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	67 %	242 tok/s	20.5 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	67 %	263 tok/s	26.3 tok/s	—	—	—	Specs →
63	Trainium	8	67 %	263 tok/s	26.3 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	67 %	162 tok/s	39.4 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	67 %	~232 tok/s	19.5 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	88 %	137 tok/s	14.7 tok/s	$3.28	$6.67	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	65 %	158 tok/s	27.8 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	90 %	251 tok/s	32.3 tok/s	$7.12	$7.87	3.6 kW	Rent →
69	A30 24GB	8	90 %	246 tok/s	29.9 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	90 %	~185 tok/s	13.9 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	88 %	118 tok/s	13.0 tok/s	$2.00	$4.70	1.2 kW	Rent →
72	A40 48GB	4	88 %	112 tok/s	11.8 tok/s	$2.36	$5.85	1.2 kW	Rent →
73	H20 141GB HBM3e	2	59 %	77.5 tok/s	35.9 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	2	87 %	77.5 tok/s	35.9 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	90 %	~207 tok/s	21.6 tok/s	$4.56	$6.12	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	67 %	238 tok/s	38.5 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	88 %	89.1 tok/s	6.9 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	87 %	47.0 tok/s	3.7 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	90 %	~203 tok/s	24.6 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	90 %	182 tok/s	18.5 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	67 %	213 tok/s	36.4 tok/s	$7.04	$9.17	2.0 kW	Rent →
82	ATOM-Max	4	65 %	109 tok/s	17.4 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	90 %	178 tok/s	19.3 tok/s	$8.81	$13.72	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	88 %	102 tok/s	14.7 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	88 %	102 tok/s	14.7 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	90 %	197 tok/s	30.8 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	90 %	136 tok/s	9.6 tok/s	$4.72	$9.65	0.6 kW	Rent →
88	L20 48GB	4	88 %	100 tok/s	14.7 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	67 %	153 tok/s	20.5 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	67 %	~150 tok/s	19.5 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	90 %	~139 tok/s	14.6 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	90 %	119 tok/s	9.6 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	90 %	~135 tok/s	13.9 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	88 %	78.2 tok/s	10.4 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	88 %	79.8 tok/s	14.7 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	67 %	139 tok/s	18.5 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	90 %	124 tok/s	30.0 tok/s	$4.00	$8.95	2.8 kW	Rent →
98	A10G 24GB	8	90 %	114 tok/s	19.3 tok/s	$8.05	$19.52	2.4 kW	Rent →
99	Atlas 300V Pro	4	88 %	44.6 tok/s	3.5 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	90 %	84.3 tok/s	6.6 tok/s	—	—	—	Specs →
101	Atlas 300V	8	90 %	68.0 tok/s	6.6 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	90 %	74.6 tok/s	32.3 tok/s	$3.68	$13.71	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	67 %	63.1 tok/s	16.4 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	90 %	52.2 tok/s	24.6 tok/s	$2.16	$11.50	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	76 %	49.6 tok/s	14.4 tok/s	—	—	1.3 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 167 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 261 tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 167 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 872 tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 261 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 1.0k tok/s

Huawei 10 parts fit

Fewest GPUs: 2 x Ascend 910C, 302 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 180 tok/s for $2.20/hr

Fewest GPUs: 2 x Gaudi 3 128GB, 499 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 423 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 263 tok/s

Other 7 parts fit

Fewest GPUs: 2 x Cloud AI 100 Ultra, 95.6 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	500 x TP2	0	198k tok/s	$2,480	700 kW	$3.48
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	46k tok/s	$1,050	—	$6.36

Size OLMo 3.1 32B Instruct 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 OLMo 3.1 32B Instruct need?

The weights take 64.5 GB at the native BF16 precision (64.5 GB at BF16, 32.2 GB at FP8, 18.1 GB at INT4). Each concurrent stream adds 256 KB of KV cache per token: 1.1 GB at 4,096 tokens and 3.0 GB at 32,768 tokens.

What is the bare minimum to run OLMo 3.1 32B Instruct?

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 OLMo 3.1 32B Instruct?

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 4 H100s, and 1,000 tokens per second takes 8 H100s across 2 replicas.

How much does OLMo 3.1 32B Instruct cost per million tokens on an H100?

About $3.05 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 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.67.

What precision does OLMo 3.1 32B Instruct 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. Fully open (weights, data and code). 48 of 64 layers attend over a 4,096-token sliding window.