NVIDIAAda Lovelace

Best AI Models for NVIDIA L4 (24.0GB)

VRAM:24.0 GB GDDR6·Bandwidth:300.0 GB/s·CUDA Cores:7,424·TDP:72W

24 GB is the enthusiast tier for running AI models locally. It comfortably handles 7B–13B models at high quality and opens the door to larger 30B models at moderate quantization.

This is one of the most popular memory tiers for local AI, found in GPUs like the RTX 4090 and RTX 3090. You can run Llama 3 8B, Mistral 7B, and Qwen 2.5 7B at Q5_K_M or Q6_K quality with fast token generation and generous context windows. Larger 14B models like DeepSeek R1 Distill fit comfortably at Q4_K_M. For even bigger models, 30B class runs at Q2–Q3, but 70B models are generally too heavy for single-GPU inference at this tier.

Runs Well

  • 7B models (Llama 3 8B, Mistral 7B) at Q5–Q8 quality
  • 13B–14B models at Q4–Q5 quality
  • Small models (3B–4B) at FP16 precision
  • Multimodal models like LLaVA 7B

Challenging

  • 30B models only at Q2–Q3 quantization
  • 70B models do not fit in VRAM
  • Large context windows with 14B+ models

What LLMs Can NVIDIA L4 Run?

123 models · 20 excellent · 38 good

Showing compatibility for NVIDIA L4

LLM models compatible with NVIDIA L4 — ranked by performance
ModelVRAMGrade
Q4_K_M·13.1 t/s tok/s·41K ctx·DECENT
14.9 GBB52
Qwen3 32B32.8B
Q4_K_M·9.6 t/s tok/s·41K ctx·BARELY RUNS
20.3 GBC41
FP16·12.7 t/s tok/s·2K ctx·DECENT
15.4 GBB51
Q4_K_M·9.5 t/s tok/s·33K ctx·BARELY RUNS
20.5 GBC41
Q4_K_M·9.5 t/s tok/s·131K ctx·BARELY RUNS
20.5 GBC41
QwQ 32B32.8B
Q4_K_M·9.5 t/s tok/s·41K ctx·BARELY RUNS
20.5 GBC41
Q4_K_M·13.2 t/s tok/s·4K ctx·DECENT
14.8 GBB52
Qwen1.5 32B32.5B
Q4_K_M·9.6 t/s tok/s·33K ctx·BARELY RUNS
20.3 GBC41
Gemma 4 31B IT32.7B
Q4_K_M·9.2 t/s tok/s·262K ctx·BARELY RUNS
21.2 GBC36
Q4_K_M·9.6 t/s tok/s·33K ctx·BARELY RUNS
20.3 GBC41
Qwen2.5 Coder 32B32.8B
Q4_K_M·9.5 t/s tok/s·33K ctx·BARELY RUNS
20.5 GBC41
Phi 414.7B
Q4_K_M·20.5 t/s tok/s·16K ctx·DECENT
9.5 GBB60
Q4_K_M·23.7 t/s tok/s·262K ctx·DECENT
8.2 GBB64
Qwen1.5 14B14.2B
Q4_K_M·18.6 t/s tok/s·33K ctx·DECENT
10.5 GBB58
Gemma 3 12B IT12.2B
Q4_K_M·24.3 t/s tok/s·33K ctx·RUNS WELL
8.0 GBA65
Qwen AgentWorld 35B A3B34.7B
Q4_K_M·9.2 t/s tok/s·262K ctx·BARELY RUNS
21.2 GBC36

NVIDIA L4 Specifications

Brand
NVIDIA
Architecture
Ada Lovelace
Compute Capability
8.9 (CUDA SM version)
VRAM
24.0 GB GDDR6
Memory Bandwidth
300.0 GB/s
CUDA Cores
7,424
Tensor Cores
240
FP16 Performance
121.00 TFLOPS
TDP
72W
Release Date
2023-03-21

Get Started

Ollama (Recommended)

$curl -fsSL https://ollama.com/install.sh | sh
$ollama run llama3:8b

LM Studio

LM Studio

Download LM Studio, search for a model, and run it with one click.

Prompt Processing

Estimated for Gemma4 12B Mtp Assistant, the compute-bound phase that reads your prompt before the first reply token appears.

2,268.8tok/s prefill

Short chat

226 ms

512 tok prompt

Long chat

1.8 s

4,096 tok prompt

Document / codebase

14.4 s

32,768 tok prompt

Prefill is compute-bound and a different number from the decode tok/s shown elsewhere on this page — how prompt processing works →

Efficiency & Value

Based on Gemma4 12B Mtp Assistant at ~24.6 tok/s decode.

Tokens per watt

0.34tok/s per W

Higher is better.

How efficiency & value are calculated →

Performance figures are estimates calibrated as of 2026-07-30 see calibration basis →

GPUs to Consider Over NVIDIA L4

Similar GPUs and upgrades with more VRAM or higher bandwidth for AI

Frequently Asked Questions

Can NVIDIA L4 run Qwen3.6 27B?

Yes, the NVIDIA L4 with 24 GB can run Qwen3.6 27B, Gemma 3 27B IT, Gemma 4 26B A4B IT, and 1520 other models. 485 models run at excellent quality, and 498 at good quality. Check the compatibility table above for the full list with VRAM usage and estimated speed.

Is NVIDIA L4 good for AI?

The NVIDIA L4 has 24 GB of GDDR6, making it excellent for running local AI models. It supports 983 models at good quality or better. With 300.0 GB/s memory bandwidth, it delivers reasonable token generation speeds. This is an enthusiast-grade GPU that handles most popular open-source LLMs.

How many parameters can NVIDIA L4 handle?

With 24 GB, the NVIDIA L4 supports models from 3B to 30B parameters depending on quantization level. At Q4_K_M (the recommended sweet spot), you can fit roughly 40B parameters. This means 7B models at high quality (Q6/Q8) or 30B+ models at Q4.

What quantization should I use on NVIDIA L4?

For the best balance of quality and speed on the NVIDIA L4, start with Q4_K_M — it preserves ~85% of the original model quality while keeping VRAM usage reasonable. With 24+ GB, you have the headroom to run 7B models at Q5_K_M or even Q6_K for noticeably better output quality. For larger 30B models, Q4_K_M remains the sweet spot.

How fast is NVIDIA L4 for AI inference?

With 300.0 GB/s memory bandwidth, the NVIDIA L4 achieves approximately 43 tokens/sec on a 7B model at Q4_K_M — that's very fast, well above conversational speed. A 14B model runs at ~22 tok/s. Token generation speed scales inversely with model size — smaller models are significantly faster.

tok/s = (300 GB/s ÷ model GB) × efficiency

Smaller models = faster inference. Memory bandwidth is the main bottleneck for token generation speed.

Estimated speed on NVIDIA L4

Real-world results typically within ±20%. Speed depends on quantization kernel, batch size, and software stack.

Learn more about tok/s estimation →

What's the best model for NVIDIA L4?

The top-rated models for the NVIDIA L4 are Qwen3.6 27B, Gemma 3 27B IT, Gemma 4 26B A4B IT. The best choice depends on your use case: coding assistants benefit from code-tuned models, while general chat works well with instruction-tuned models like Llama or Qwen.

What power supply and cooling does NVIDIA L4 need?

The NVIDIA L4 has a TDP of 72 W. A good rule of thumb is to provide at least double the GPU's TDP to cover the rest of the system — that means a 550 W PSU or larger. It's a relatively low-power card, so most mid-tower cases with basic airflow handle it comfortably. Still, ensure the GPU slot has clearance and vents aren't obstructed during long inference sessions.