Mistral AI Mistral Small

Mistral Small 3.2 24B hardware requirements

Fits on one Instinct MI350X OAM at BF16. For 1,000 tokens/s at 32k context you need 4 x H100 SXM5 80GB at $7.16/hr or 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00/hr.

Parameters

24.0B

Active per token

All

dense model

KV cache per token

160 KB

16-bit cache

Native precision

BF16

Rent the cheapest fit: 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition from $0.50/hr Model card

Bare minimum

Fewest GPUs

1 x Instinct MI350X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 123
Per hour: $6.16
Per million tokens: $13.89
Memory used: 18%

Rent on DigitalOcean

Cheapest per hour

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 27.6
Per hour: $0.50
Per million tokens: $5.03
Memory used: 53%

Rent on RunPod

Cheapest per million tokens

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 704
Per hour: $2.00
Per million tokens: $0.79
Memory used: 60%

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; 37 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

GPU	TP	Memory used	Tokens/s	Per stream	Per hour	Per M tokens	kW
Instinct MI455X OAM	1	53%	~2.5k	79.5				Specs
Rubin SXM	1	79%	2.4k	75.1				Specs
Instinct MI355X OAM	1	79%	873	27.3			1.4	Specs
GB300 NVL72 GPU 288GB	1	79%	873	27.3	$8.62	$2.74		Rent
GB200 NVL72 GPU 186GB	2	62%	1.7k	51.9	$21.00	$3.52		Rent
Instinct MI350X OAM	1	79%	873	27.3	$6.16	$1.96	1.0	Rent
B300 SXM 262GB	1	87%	873	27.3	$7.40	$2.35	1.4	Rent
B200 SXM 180GB	2	64%	1.6k	49.9	$8.18	$1.42	2.0	Rent
B100 SXM 192GB	2	60%	1.7k	51.9			1.4	Specs
Instinct MI325X OAM	1	89%	655	20.5	$3.80	$1.61	1.0	Rent
Instinct MI300X 192GB	2	60%	1.1k	34.4	$4.78	$1.21	1.5	Rent
BR100	4	90%	629	19.7			2.2	Specs
H200 SXM 141GB	2	81%	996	31.1	$5.98	$1.67	1.4	Rent
H100 SXM5 80GB	4	72%	1.3k	41.1	$7.16	$1.51	2.8	Rent
Data Center GPU Max 1550 128GB	2	90%	680	21.2			1.2	Specs
H200 NVL 141GB	2	81%	~996	31.1	$7.58	$2.11	1.2	Rent
H100 NVL 94GB	4	61%	~1.5k	47.9	$12.76	$2.31	1.6	Rent
H100 PCIe 80GB	4	72%	786	24.6	$10.00	$3.53	1.4	Rent
H800 PCIe 80GB	4	72%	786	24.6			1.4	Specs
RTX PRO 6000 Blackwell Workstation Edition	4	60%	704	22.0	$7.20	$2.84	2.4	Rent
RTX PRO 6000 Blackwell Server Edition	4	60%	~628	19.6	$2.36	$1.04	2.4	Rent
RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	60%	704	22.0	$2.00	$0.79	1.2	Rent
Data Center GPU Max 1100 48GB	8	61%	912	28.5			2.4	Specs
GeForce RTX 5090 32GB	8	92%	1.3k	41.6	$7.92	$1.65	4.6	Rent
MI250X 128GB	2	90%	680	21.2			1.1	Specs
RTX 6000 Ada 48GB	8	61%	~713	22.3	$6.72	$2.62	2.4	Rent
Instinct MI250 128GB	2	90%	680	21.2			1.0	Specs
L40S 48GB	8	61%	641	20.0	$6.96	$3.01	2.8	Rent
A800 PCIe 80GB	4	72%	761	23.8			1.2	Specs
A100 PCIe 80GB	4	72%	761	23.8	$5.40	$1.97	1.2	Rent
A100 SXM4 80GB	4	72%	801	25.0	$5.60	$1.94	1.6	Rent
A100 PCIe 40GB	8	73%	1.2k	36.1			2.0	Specs
A100 SXM4 40GB	8	73%	1.2k	36.1	$10.32	$2.48	3.2	Rent
GeForce RTX 5090 D V2 24GB 2 nodes	16	62%	~1.8k	55.0			9.2	Specs
RTX 5880 Ada 48GB	8	61%	~713	22.3			2.3	Specs
RTX 5000 Ada 32GB	8	92%	~428	13.4	$6.64	$4.31	2.0	Rent
RTX PRO 5000 Blackwell 48GB	8	61%	~998	31.2	$7.68	$2.14	2.4	Rent
RTX PRO 5000 Blackwell 72GB	4	80%	~528	16.5			1.2	Specs
RTX PRO 4500 Blackwell 32GB	8	92%	~665	20.8	$5.76	$2.41	1.6	Rent
Radeon AI PRO R9700 32GB	8	92%	475	14.8			2.4	Specs
Instinct MI100 32GB	8	92%	912	28.5			2.4	Specs
Arc Pro B70 32GB	8	92%	~451	14.1			1.8	Specs
L40 48GB	8	61%	641	20.0	$6.56	$2.84	2.4	Rent
Instinct MI210 PCIe	4	90%	644	20.1			1.2	Specs
GeForce RTX 4090 24GB 2 nodes	16	62%	1.3k	41.3	$9.60	$2.02	7.2	Rent
A30 24GB 2 nodes	16	62%	1.2k	38.2	$11.73	$2.67	2.6	Rent
RTX 4500 Ada 24GB 2 nodes	16	62%	~566	17.7			3.4	Specs
RTX A6000 48GB	8	61%	570	17.8	$4.00	$1.95	2.4	Rent
A40 48GB	8	61%	517	16.1	$3.92	$2.11	2.4	Rent
H20 141GB HBM3e	2	81%	830	25.9			0.8	Specs
H20 96GB	4	60%	1.6k	49.1			1.6	Specs
RTX PRO 4000 Blackwell 24GB 2 nodes	16	62%	~880	27.5	$9.12	$2.88	2.3	Rent
BI-V100 Tiangai 100	8	92%	891	27.8			2.0	Specs
RTX A5500 24GB 2 nodes	16	62%	~1.0k	31.4			3.7	Specs
V100S PCIe 32GB	8	92%	842	26.3	$7.04	$2.32	2.0	Rent
A10 24GB 2 nodes	16	62%	786	24.6	$20.64	$7.29	2.4	Rent
Radeon PRO W7900 Dual Slot 48GB	8	61%	641	20.0			2.4	Specs
Radeon PRO W7900 48GB	8	61%	641	20.0			2.4	Specs
Radeon RX 7900 XTX 24GB 2 nodes	16	62%	1.3k	39.3			5.7	Specs
L4 24GB 2 nodes	16	62%	393	12.3	$7.84	$5.54	1.2	Rent
L20 48GB	8	61%	641	20.0			2.2	Specs
Arc Pro B65 32GB	8	92%	~451	14.1			1.6	Specs
Arc Pro B60 24GB 2 nodes	16	62%	~597	18.7			3.2	Specs
L2 24GB 2 nodes	16	62%	393	12.3				Specs
RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	62%	~566	17.7			1.1	Specs
Radeon PRO W7800 48GB	8	61%	641	20.0			2.1	Specs
Radeon PRO W7800 32GB	8	92%	428	13.4			2.1	Specs
GeForce RTX 3090 24GB 2 nodes	16	62%	1.2k	38.3	$8.00	$1.81	5.6	Rent
A10G 24GB 2 nodes	16	62%	786	24.6			4.8	Specs
GeForce RTX 3090 Ti 24GB 2 nodes	16	62%	1.3k	41.3	$7.36	$1.55	7.2	Rent
Radeon PRO W6800 32GB	8	92%	380	11.9			2.0	Specs
RTX A5000 24GB 2 nodes	16	62%	1.0k	31.4	$4.32	$1.19	3.7	Rent
Radeon PRO V710 28GB 2 nodes	16	53%	587	18.3			2.5	Specs

AMD vs NVIDIA

At BF16 the NVIDIA pick is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 704 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 1.1k tokens/s for $4.78/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 49 parts fit

Best value: 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 704 tok/s for $2.00/hr

Fewest GPUs: 1 x Rubin SXM, 2.4k tok/s

AMD 17 parts fit

Best value: 2 x Instinct MI300X 192GB, 1.1k tok/s for $4.78/hr

Fewest GPUs: 1 x Instinct MI455X OAM, 2.5k tok/s

Intel 5 parts fit

Fewest GPUs: 2 x Data Center GPU Max 1550 128GB, 680 tok/s

Other 2 parts fit

Fewest GPUs: 4 x BR100, 629 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	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A 1,000 x H200 SXM 141GB	500 x TP2	0	498k	$2,990	700 kW	$1.67
B 100 x GB200 NVL72 GPU 186GB	50 x TP2	0	83k	$1,050	kW	$3.52

Size Mistral Small 3.2 24B 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 Mistral Small 3.2 24B need?

The weights take 48.0 GB at the native BF16 precision (48.0 GB at BF16, 24.0 GB at FP8, 13.5 GB at INT4). Each concurrent stream adds 160 KB of KV cache per token: 0.7 GB at 4,096 tokens and 5.4 GB at 32,768 tokens.

What is the bare minimum to run Mistral Small 3.2 24B?

1 x Instinct MI350X 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 Mistral Small 3.2 24B?

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 4 H100s across 1 replicas.

How much does Mistral Small 3.2 24B cost per million tokens on an H100?

About $1.51 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.79 per GPU-hour. The cheapest part per token is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.79.

What precision does Mistral Small 3.2 24B 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.