What LLMs Can the NVIDIA L4 24GB Run?

The NVIDIA L4 24GB has 24 GB of VRAM and 300 GB/s of bandwidth. It runs 41 of 58 models fully at Q4_K_M with 8K context, up to Qwen 3.5 35B-A3B, and generates about 36 tokens/s on an 8B model.

NVIDIA · Ada Lovelace · Released 2023-03 · Updated October 14, 2026

41 of 58 models fit

NVIDIA L4 24GB (24.0 GB) at Q4_K_M with 8K context. Largest that fits fully: Qwen 3.5 35B-A3B.

ModelNeedsStatusSpeedBest quant that fits
Qwen 2.5 32B21.4 GBFits~9.2 tok/sQ4_K_M
Qwen 2.5 Coder 32B21.4 GBFits~9.2 tok/sQ4_K_M
DeepSeek R1 Distill Qwen 32B21.4 GBFits~9.2 tok/sQ4_K_M
Qwen 3 32B21.4 GBFits~9.2 tok/sQ4_K_M
Qwen 3.5 35B-A3B21.0 GBFits~71 tok/sQ4_K_M
Qwen 3.6 35B-A3B21.0 GBFits~71 tok/sQ4_K_M
Gemma 4 31B19.8 GBFits~9.8 tok/sQ4_K_M
Qwen 3 30B-A3B18.8 GBFits~58 tok/sQ5_K_M
Qwen3-Coder 30B-A3B18.8 GBFits~58 tok/sQ5_K_M
Gemma 3 27B17.6 GBFits~11 tok/sQ5_K_M
Qwen 3.5 27B16.9 GBFits~12 tok/sQ6_K
Qwen 3.6 27B16.9 GBFits~12 tok/sQ6_K
Qwen 3.8 27B16.9 GBFits~12 tok/sQ6_K
Mistral Small 3.1 24B15.6 GBFits~13 tok/sQ6_K
Gemma 4 26B-A4B15.6 GBFits~61 tok/sQ6_K
gpt-oss-20b12.9 GBFits~71 tok/sMXFP4 (official)
Qwen 2.5 14B10.7 GBFits~20 tok/sQ8_0
DeepSeek R1 Distill Qwen 14B10.7 GBFits~20 tok/sQ8_0
Phi-4 14B10.6 GBFits~20 tok/sQ8_0
Qwen 3 14B10.5 GBFits~20 tok/sQ8_0
Ministral 3 14B9.7 GBFits~22 tok/sQ8_0
Mistral Nemo 12B9.0 GBFits~24 tok/sQ8_0
Gemma 3 12B8.8 GBFits~24 tok/sQ8_0
Ministral 3 8B6.7 GBFits~34 tok/sFP16 / BF16
Qwen 3 8B6.6 GBFits~35 tok/sFP16 / BF16
Qwen 3.5 9B6.4 GBFits~34 tok/sFP16 / BF16
Llama 3.1 8B6.3 GBFits~36 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 8B6.3 GBFits~36 tok/sFP16 / BF16
Mistral 7B v0.35.7 GBFits~39 tok/sFP16 / BF16
Gemma 4 E4B5.7 GBFits~39 tok/sFP16 / BF16
Qwen 2.5 7B5.6 GBFits~40 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 7B5.6 GBFits~40 tok/sFP16 / BF16
Qwen 3 4B4.2 GBFits~63 tok/sFP16 / BF16
Phi-4 Mini4.1 GBFits~67 tok/sFP16 / BF16
Gemma 4 E2B4.0 GBFits~61 tok/sFP16 / BF16
Gemma 3 4B3.7 GBFits~68 tok/sFP16 / BF16
Qwen 3.5 4B3.7 GBFits~70 tok/sFP16 / BF16
Ministral 3 3B3.5 GBFits~76 tok/sFP16 / BF16
Llama 3.2 3B3.5 GBFits~79 tok/sFP16 / BF16
Llama 3.2 1B1.7 GBFits~213 tok/sFP16 / BF16
Gemma 3 1B1.6 GBFits~285 tok/sFP16 / BF16
Qwen 3.5 122B-A10B70.2 GB32% on GPU~8.5 tok/s–
Mistral Small 4 119B68.2 GB33% on GPU~13 tok/s–
gpt-oss-120b67.2 GB34% on GPU~17 tok/s–
Llama 4 Scout64.0 GB35% on GPU~5.1 tok/s–
GLM-4.5-Air62.1 GB36% on GPU~7.1 tok/s–
Qwen3-Next 80B-A3B46.3 GB49% on GPU~29 tok/s–
Qwen 2.5 72B44.5 GB51% on GPU~1.5 tok/s–
Llama 3.3 70B43.2 GB52% on GPU~1.6 tok/s–
DeepSeek R1 Distill Llama 70B43.2 GB52% on GPU~1.6 tok/s–
Mixtral 8x7B28.0 GB81% on GPU~13 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––
Qwen 3 235B-A22B135 GBToo small––
MiniMax-M2133 GBToo small––

NVIDIA L4 24GB for local AI

A low-power 24 GB cloud GPU common on Google Cloud. It holds 27–32B models at Q4_K_M and supports FP8, but 300 GB/s of bandwidth makes it slower per token than consumer 24 GB cards.

Among the 8 devices we track with about 24 GB of usable memory, the NVIDIA L4 24GB is the slowest at generating text, because generation speed follows memory bandwidth. Others in the same class: RTX 3090 Ti 24GB, RTX 4090 24GB, Radeon RX 7900 XTX 24GB, RTX 3090 24GB. They all hold the same models; the faster ones simply answer sooner.

Our picks for the NVIDIA L4 24GB

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-rounderQwen 3.6 27BQ6_K98K~12 tok/s
CodingQwen 3.6 27BQ6_K98K~12 tok/s
Fastest repliesQwen 3.6 35B-A3BQ4_K_M88K~71 tok/s
Long documents (64K+)Qwen 3.8 27BQ6_K98K~12 tok/s
Images in, text outGemma 4 31BQ4_K_M43K~9.8 tok/s
Image generationQwen-ImageFP8––
SpecValue
Memory24 GB
Memory bandwidth300 GB/s
ArchitectureAda Lovelace
Released2023-03
Llama 3.1 8B Q4_K_M speed~36 tok/s
Largest model that fitsQwen 3.5 35B-A3B

Image and video models on the NVIDIA L4 24GB

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 GB73% on GPU23.6 GBTight fit
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 40GB with 40 GB and 418% more bandwidth.

NVIDIA L4 24GB for LLMs: common questions

Can the NVIDIA L4 24GB run Llama 3.3 70B?

Only partly. It needs about 43.2 GB at Q4_K_M, so about 52% fits in VRAM and the rest runs from system RAM, which is slow.

Can the NVIDIA L4 24GB 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 Q4_K_M.

Can the NVIDIA L4 24GB 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 Q8_0.

How fast is the NVIDIA L4 24GB for LLMs?

About 36 tokens per second for an 8B model at Q4_K_M, based on its 300 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.

Related