NVIDIA / discrete

NVIDIA L40S

48 GB decides what fits. 864 GB/s decides how fast it runs once it does.

Open in the calculator →Best models for 48 GB, on every card that size →

Quick answer

The NVIDIA L40S fits 226 of 320 sized models entirely in its 48 GB; the largest widely used one is Kimi-Linear-48B-A3B-Instruct (30.4 GB at Q4_K_M). Memory decides what fits; its 864 GB/s of bandwidth decides how fast it answers.

Largest popular model that fitsKimi-Linear-48B-A3B-Instruct49B params · Q4_K_M · needs 30.4 GB~314 tok/s Faster than you readOpen in the calculator →
Most downloaded that fitsQwen3.8-27B28B params · Q4_K_M · needs 18.6 GB~40 tok/s Faster than you readOpen in the calculator →

At 8,192 tokens of context with the whole model in device memory. Speeds are estimates from memory bandwidth, not benchmarks run on this card.

NVIDIA L40S · 48 GB · 864 GB/s · Q4_K_M where published · 8,192 tokensSpecificationpublished dataDecode speedestimate

226 of 320 fit entirely

226 of the 320 models the engine can size fit entirely in device memory at Q4_K_M where it is published, otherwise the nearest published format, and 8,192 tokens. 18 run with some layers on system memory (32 GB assumed), and 76 do not run at all.

71%of the models the engine can size fit entirely

48GB of device memory

864GB/s memory bandwidth

18run with system memory

76do not run at 8,192 tokens

Featured models · Q4_K_M at 8,192 tokens, including offload

Model	Needs	Verdict	Decode	Calculator
Qwen3.5-2B2.3B parameters	2.32 GB	fits	~476 tok/s	Open →
NVIDIA-Nemotron-3-Nano-4B-BF164.0B parameters	3.46 GB	fits	~221 tok/s	Open →
Qwen3-VL-8B-Instruct8.8B parameters	7.04 GB	fits	~100 tok/s	Open →
Qwen3.5-9B9.7B parameters	7.05 GB	fits	~121 tok/s	Open →
gemma-4-12B-it12B parameters	9.54 GB	fits	~72 tok/s	Open →
gpt-oss-20b21B parameters	13.8 GB	fits	~235 tok/s	Open →
Qwen3.8-27B28B parameters	18.6 GB	fits	~40 tok/s	Open →
Qwen3.6-27B28B parameters	18.6 GB	fits	~40 tok/s	Open →
Kimi-Linear-48B-A3B-Instruct49B parameters	30.4 GB	fits	~314 tok/s	Open →
Qwen3.8-Flash-Next180B parameters	111.6 GB	needs 64.5 GB of system RAM; 32 GB assumed	does not run	Open →
DeepSeek-V4-Flash-Vision-Exp305B parameters	188.1 GB	needs 143.7 GB of system RAM; 32 GB assumed	does not run	Open →
GLM-5.3-Flash321B parameters	198.3 GB	needs 153.4 GB of system RAM; 32 GB assumed	does not run	Open →

These examples are selected from prominent labs using the catalogue’s latest Hugging Face 30-day downloads and repository-creation freshness signal, with newer releases guaranteed a place. Offloaded rows assume 32 GB of system RAM, and a speed is only shown for a row that runs. Fitting in memory is not the same as loading: whether the runtime and version you have supports each architecture and format on this machine has not been tested here. Each “Open” link carries the same model, format, context and RAM into the calculator.

Fits entirely in NVIDIA L40S memory — 226 of 320 sized models

Every model that fitswhole model resident · offload off

The most demanding model that fits is Qwen2.5-VL-72B-Instruct at Q4_K_M: 47.3 GB of the 48 GB, leaving 0.70 GB spare.

Fully resident at 8,192 tokens (or the model’s own maximum, where that is shorter), offload off, at 864 GB/s. Each row is a run of the engine for this configuration; the rows start with current, prominent releases and “fits” is memory, not a tested runtime. Older or less prominent models remain available through this search and “Show all”.

Find a model in this list

Showing 40 of 226 models that fit.

Model	Format	Needs	Spare	Decode	Downloads / 30d	Calculator
Qwen3.8-27BQwen · 28B params	Q4_K_M	18.6 GB	29.4 GB	~40 tok/s	6.9M	Open →
gemma-4-26B-A4B-itGoogle · 26B params	Q4_K_M	16.8 GB	31.2 GB	~221 tok/s	13M	Open →
gemma-4-31B-itGoogle · 31B params	Q4_K_M	22.2 GB	25.8 GB	~29 tok/sAbout reading pace	9.9M	Open →
Qwen3.5-9BQwen · 9.7B params	Q4_K_M	7.05 GB	41.0 GB	~121 tok/s	9M	Open →
Qwen3.5-4BQwen · 4.7B params	Q4_K_M	3.98 GB	44.0 GB	~218 tok/s	7.8M	Open →
Qwen3.6-35B-A3BQwen · 36B params	Q4_K_M	23.1 GB	24.9 GB	~317 tok/s	3.3M	Open →
gemma-4-12B-itGoogle · 12B params	Q4_K_M	9.54 GB	38.5 GB	~72 tok/s	1.9M	Open →
Qwen3.6-27BQwen · 28B params	Q4_K_M	18.6 GB	29.4 GB	~40 tok/s	2.5M	Open →
Qwen3.5-2BQwen · 2.3B params	Q4_K_M	2.32 GB	45.7 GB	~476 tok/s	4.9M	Open →
gemma-4-E4B-itGoogle · 8.0B params	Q4_K_M	5.86 GB	42.1 GB	~194 tok/s	4.4M	Open →
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16NVIDIA · 32B params	Q4_K_M	20.3 GB	27.7 GB	~30 tok/s	530K	Open →
NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA · 4.0B params	Q4_K_M	3.46 GB	44.5 GB	~221 tok/s	3.4M	Open →
gemma-4-E2B-itGoogle · 5.1B params	Q4_K_M	4.02 GB	44.0 GB	~513 tok/s	3M	Open →
Qwen3.5-0.8BQwen · 873M params	Q4_K_M	1.46 GB	46.5 GB	~1043 tok/s	2.6M	Open →
Qwen3-VL-8B-InstructQwen · 8.8B params	Q4_K_M	7.04 GB	41.0 GB	~100 tok/s	14.6M	Open →
Qwen3.5-27BQwen · 28B params	Q4_K_M	18.6 GB	29.4 GB	~40 tok/s	1.9M	Open →
North-Micro-Vision-InstructCohereLabs · 2.5B params	Q4_K_M	2.99 GB	45.0 GB	~313 tok/s	180.7K	Open →
Qwen3.5-35B-A3BQwen · 36B params	Q4_K_M	23.1 GB	24.9 GB	~317 tok/s	1.6M	Open →
granite-4.2-8bIBM · 8.8B params	Q4_K_M	7.49 GB	40.5 GB	~91 tok/s	128.4K	Open →
GLM-4.7-Flashzai-org · 31B params	Q4_K_M	20.4 GB	27.6 GB	~30 tok/sAbout reading pace	1.8M	Open →
granite-4.1-3bIBM · 3.4B params	Q4_K_M	3.72 GB	44.3 GB	~212 tok/s	521.8K	Open →
LFM2.5-2.6BLiquidAI · 2.7B params	Q4_K_M	2.61 GB	45.4 GB	~340 tok/s	108.1K	Open →
Qwen3-VL-4B-InstructQwen · 4.4B params	Q4_K_M	4.72 GB	43.3 GB	~159 tok/s	3.5M	Open →
granite-4.1-30bIBM · 29B params	Q4_K_M	20.4 GB	27.6 GB	~30 tok/sAbout reading pace	301.5K	Open →
Qwen3-VL-2B-InstructQwen · 2.1B params	Q4_K_M	3.05 GB	45.0 GB	~288 tok/s	2.8M	Open →
granite-4.2-3bIBM · 3.7B params	Q4_K_M	3.72 GB	44.3 GB	~212 tok/s	50.6K	Open →
LFM2.5-230MLiquidAI · 230M params	Q4_K_M	1.05 GB	46.9 GB	~2371 tok/s	88.6K	Open →
Qwen3-0.6BQwen · 752M params	Q4_K_M	2.22 GB	45.8 GB	~441 tok/s	29.7M	Open →
gpt-oss-20bOpenAI · 21B params	Q4_K_M	13.8 GB	34.2 GB	~235 tok/s	6.6M	Open →
granite-4.2-30bIBM · 29B params	Q4_K_M	20.7 GB	27.3 GB	~30 tok/sAbout reading pace	35.8K	Open →
granite-4.1-8bIBM · 8.8B params	Q4_K_M	7.49 GB	40.5 GB	~91 tok/s	179.3K	Open →
NVIDIA-Nemotron-3-Nano-30B-A3B-BF16NVIDIA · 32B params	Q4_K_M	20.3 GB	27.7 GB	~30 tok/s	875.2K	Open →
LFM2.5-VL-3BLiquidAI · 3.1B params	Q4_K_M	2.85 GB	45.1 GB	~340 tok/s	26.7K	Open →
MiMo-V2.6-Distill-Qwen-9BNEWXiaomiMiMo · 9.4B params	Q4_K_M	6.90 GB	41.1 GB	~121 tok/s	14.6K	Open →
Qwen3-4B-Instruct-2507Qwen · 4.0B params	Q4_K_M	4.72 GB	43.3 GB	~159 tok/s	3.7M	Open →
Ling-3.0-tinyinclusionAI · 7.9B params	Q4_K_M	5.85 GB	42.1 GB	~685 tok/s	17.7K	Open →
Qwen3-8BQwen · 8.2B params	Q4_K_M	7.04 GB	41.0 GB	~100 tok/s	10.7M	Open →
Olmo-3-7B-Instructallenai · 7.3B params	Q4_K_M	7.96 GB	40.0 GB	~85 tok/s	481.7K	Open →
Qwen3-4BQwen · 4.0B params	Q4_K_M	4.51 GB	43.5 GB	~159 tok/s	7.8M	Open →
LFM2.5-8B-A1BLiquidAI · 8.5B params	Q4_K_M	6.06 GB	41.9 GB	~530 tok/s	32.4K	Open →

Sized at the format most people actually download, not at FP16. The catalogue ordered by downloads →

Rent it instead

Rent an L40S by the hour

What the NVIDIA L40S costs on Vast.ai and RunPod right now, per machine, from one card to eight. Every price is read from the provider's own API.

Prices · 18:30 UTC, 22 Sept

1× L40S · 48 GB

$0.69per hour

Vast.ai · Marketplace · 98.5% reliableRunPod $0.79/h Community CloudRent on Vast.ai

2× L40S · 96 GB

$1.60per hour

Vast.ai · Marketplace · 99.5% reliableRent on Vast.ai

4× L40S · 192 GB

$1.87per hour

Vast.ai · Marketplace · 99.2% reliableRent on Vast.ai

8× L40S · 384 GB

$3.73per hour

Vast.ai · Marketplace · 99.9% reliableRent on Vast.ai

Buy one or rent one?

Your price, your hours, your electricity. Everything else is arithmetic.

Price you would pay ($)

Resale after 3 years ($)

Hours of use per day4 h

$ per kWh

Watts

Rent $/h

Watts start at the NVIDIA L40S’s published board power (350 W), an upper bound: decoding rarely holds a card at its limit. Rent starts at the cheapest live price (Vast.ai). The electricity price is a placeholder — put yours in.

At 4 h a day, renting costs $1,007.40 a year.

That is $83.95 a month, and nothing when the machine is stopped.
Owning would cost $127.75 a year in electricity, on top of the price.
Every $1,000 of purchase price takes 1.1 years of this use to earn back.

Enter the price you would pay to see the exact break-even.

On-demand prices for the whole machine. Vast hosts below 98% measured reliability are left out; RunPod’s Community Cloud is vetted third-party hosts and its Secure Cloud is data-centre capacity. Every rentable card compared.

Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank

Under the hood

The specification behind every figure

What the manufacturer publishes for this device, and the pages it was read from.

Manufacturer specification

Memory scope	dedicated
Capacity	48 GB
Published options	48 GB
Bandwidth	864 GB/s
Memory type	GDDR6
Bus width	Not published
FP32 peak	Not published
Dense matrix peak	Not published without sparsity
Max power consumption	350 W

Source ledger

NVIDIA L40S GPU ↗

Retrieved 2026-09-14

Caveats

The page's Tensor Core throughput figures are footnoted "With Sparsity" and are not used.
A passively cooled data-center card: the power figure assumes server airflow.
No dense matrix throughput is published for this device in a form this catalogue accepts, so no compute roof is priced and time to first token is withheld.

Published capacity is a hardware ceiling, not guaranteed free runtime memory. The calculator shows the runtime reserve separately rather than folding it into a single number.

Run the diagnostic on the NVIDIA L40S →