What LLMs Can the L40S 48GB Run?

The L40S 48GB has 48 GB of VRAM and 864 GB/s of bandwidth. It runs 46 of 58 models fully at Q4_K_M with 8K context, up to Qwen3-Next 80B-A3B, and generates about 101 tokens/s on an 8B model.

NVIDIA · Ada Lovelace · Released 2023-08 · Updated October 14, 2026

46 of 58 models fit

L40S 48GB (48.0 GB) at Q4_K_M with 8K context. Largest that fits fully: Qwen3-Next 80B-A3B.

ModelNeedsStatusSpeedBest quant that fits
Qwen 2.5 72B44.5 GBFits~12 tok/sQ4_K_M
Llama 3.3 70B43.2 GBFits~13 tok/sQ4_K_M
DeepSeek R1 Distill Llama 70B43.2 GBFits~13 tok/sQ4_K_M
Mixtral 8x7B28.0 GBFits~55 tok/sQ6_K
Qwen 2.5 32B21.4 GBFits~26 tok/sQ8_0
Qwen 2.5 Coder 32B21.4 GBFits~26 tok/sQ8_0
DeepSeek R1 Distill Qwen 32B21.4 GBFits~26 tok/sQ8_0
Qwen 3 32B21.4 GBFits~26 tok/sQ8_0
Qwen 3.5 35B-A3B21.0 GBFits~133 tok/sQ8_0
Qwen 3.6 35B-A3B21.0 GBFits~133 tok/sQ8_0
Gemma 4 31B19.8 GBFits~28 tok/sQ8_0
Qwen 3 30B-A3B18.8 GBFits~110 tok/sQ8_0
Qwen3-Coder 30B-A3B18.8 GBFits~110 tok/sQ8_0
Gemma 3 27B17.6 GBFits~32 tok/sQ8_0
Qwen 3.5 27B16.9 GBFits~33 tok/sQ8_0
Qwen 3.6 27B16.9 GBFits~33 tok/sQ8_0
Qwen 3.8 27B16.9 GBFits~33 tok/sQ8_0
Mistral Small 3.1 24B15.6 GBFits~36 tok/sQ8_0
Gemma 4 26B-A4B15.6 GBFits~131 tok/sQ8_0
gpt-oss-20b12.9 GBFits~155 tok/sMXFP4 (official)
Qwen 2.5 14B10.7 GBFits~56 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 14B10.7 GBFits~56 tok/sFP16 / BF16
Phi-4 14B10.6 GBFits~56 tok/sFP16 / BF16
Qwen 3 14B10.5 GBFits~57 tok/sFP16 / BF16
Ministral 3 14B9.7 GBFits~62 tok/sFP16 / BF16
Mistral Nemo 12B9.0 GBFits~68 tok/sFP16 / BF16
Gemma 3 12B8.8 GBFits~68 tok/sFP16 / BF16
Ministral 3 8B6.7 GBFits~95 tok/sFP16 / BF16
Qwen 3 8B6.6 GBFits~97 tok/sFP16 / BF16
Qwen 3.5 9B6.4 GBFits~97 tok/sFP16 / BF16
Llama 3.1 8B6.3 GBFits~101 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 8B6.3 GBFits~101 tok/sFP16 / BF16
Mistral 7B v0.35.7 GBFits~110 tok/sFP16 / BF16
Gemma 4 E4B5.7 GBFits~109 tok/sFP16 / BF16
Qwen 2.5 7B5.6 GBFits~112 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 7B5.6 GBFits~112 tok/sFP16 / BF16
Qwen 3 4B4.2 GBFits~173 tok/sFP16 / BF16
Phi-4 Mini4.1 GBFits~184 tok/sFP16 / BF16
Gemma 4 E2B4.0 GBFits~169 tok/sFP16 / BF16
Gemma 3 4B3.7 GBFits~187 tok/sFP16 / BF16
Qwen 3.5 4B3.7 GBFits~193 tok/sFP16 / BF16
Ministral 3 3B3.5 GBFits~211 tok/sFP16 / BF16
Llama 3.2 3B3.5 GBFits~218 tok/sFP16 / BF16
Llama 3.2 1B1.7 GBFits~576 tok/sFP16 / BF16
Gemma 3 1B1.6 GBFits~719 tok/sFP16 / BF16
Qwen3-Next 80B-A3B46.3 GBTight fit~120 tok/sIQ4_XS
Qwen 3 235B-A22B135 GB33% on GPU~4.1 tok/s–
MiniMax-M2133 GB34% on GPU~8.2 tok/s–
Qwen 3.5 122B-A10B70.2 GB64% on GPU~15 tok/sQ2_K
Mistral Small 4 119B68.2 GB66% on GPU~24 tok/sQ2_K
gpt-oss-120b67.2 GB67% on GPU~30 tok/sQ2_K
Llama 4 Scout64.0 GB70% on GPU~10 tok/sQ2_K
GLM-4.5-Air62.1 GB73% on GPU~15 tok/sQ2_K
Kimi K2.6566 GBToo small––
DeepSeek R1 671B380 GBToo small––
Llama 3.1 405B234 GBToo small––
Qwen 3.5 397B-A17B226 GBToo small––
DeepSeek V4 Flash161 GBToo small––

L40S 48GB for local AI

A common cloud inference card. Same memory as the RTX 6000 Ada, with passive cooling for servers.

Among the 6 devices we track with about 48 GB of usable memory, the L40S 48GB ranks 2 of 6 for text-generation speed, because generation speed follows memory bandwidth. Others in the same class: RTX 6000 Ada 48GB, RTX A6000 48GB, Apple M4 Max 64GB, Apple M1 Max 64GB. They all hold the same models; the faster ones simply answer sooner.

Our picks for the L40S 48GB

The model we would run first for each job, chosen from our recommendations by what fits fully with an 8K context. "Best quant" is the highest quality that fits; "max context" is at Q4_K_M or the native format.

TaskModelBest quantMax contextSpeed
Best all-rounderLlama 3.3 70BQ4_K_M14K~13 tok/s
CodingQwen 3.6 27BQ8_0256K~33 tok/s
Fastest repliesQwen 3.6 35B-A3BQ8_0256K~133 tok/s
Long documents (64K+)Qwen 3.8 27BQ8_0256K~33 tok/s
Images in, text outGemma 4 31BQ8_0256K~28 tok/s
Image generationFLUX.2 [dev]FP8––
SpecValue
Memory48 GB
Memory bandwidth864 GB/s
ArchitectureAda Lovelace
Released2023-08
Llama 3.1 8B Q4_K_M speed~101 tok/s
Largest model that fitsQwen3-Next 80B-A3B

Image and video models on the L40S 48GB

Peak VRAM with model swapping (ComfyUI default on smaller cards).

ModelFP8FP8 fitGGUF Q4Q4 fit
Stable Diffusion 1.51.8 GBFits1.7 GBFits
SDXL 1.04.1 GBFits4.1 GBFits
Stable Diffusion 3.5 Medium6.4 GBFits6.4 GBFits
Stable Diffusion 3.5 Large8.6 GBFits6.4 GBFits
FLUX.1 [dev]12.1 GBFits7.3 GBFits
FLUX.1 [schnell]12.1 GBFits7.3 GBFits
Qwen-Image20.4 GBFits12.3 GBFits
Wan 2.2 A14B (T2V / I2V)18.8 GBFits13.2 GBFits
Wan 2.2 TI2V 5B8.0 GBFits6.5 GBFits
HunyuanVideo18.3 GBFits13.1 GBFits
FLUX.1 Kontext [dev]12.3 GBFits7.5 GBFits
FLUX.2 [dev]31.0 GBFits23.6 GBFits
Z-Image Turbo6.8 GBFits4.9 GBFits
Qwen-Image-Edit 250920.5 GBFits12.4 GBFits
LTX-219.2 GBFits12.6 GBFits

Upgrade path

The next step up is the A100 80GB with 80 GB and 136% more bandwidth.

L40S 48GB for LLMs: common questions

Can the L40S 48GB run Llama 3.3 70B?

Yes. Llama 3.3 70B needs about 43.2 GB at Q4_K_M with 8K context, and the best quality that fits is Q4_K_M.

Can the L40S 48GB run Qwen 3 32B?

Yes. Qwen 3 32B needs about 21.4 GB at Q4_K_M with 8K context, and the best quality that fits is Q8_0.

Can the L40S 48GB run Qwen 3 14B?

Yes. Qwen 3 14B needs about 10.5 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

How fast is the L40S 48GB for LLMs?

About 101 tokens per second for an 8B model at Q4_K_M, based on its 864 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.

Related