Quick Answer: NVIDIA L4 offers 24GB VRAM and starts around current market pricing. It delivers approximately 79 tokens/sec on context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16. It typically draws 72W under load.
This GPU offers reliable throughput for local AI workloads. Pair it with the right model quantization to hit your desired tokens/sec, and monitor prices below to catch the best deal.
Buy directly on Amazon with fast shipping and reliable customer service.
💡 Not ready to buy? Try cloud GPUs first
Test NVIDIA L4 performance in the cloud before investing in hardware. Pay by the hour with no commitment.
| Model | Quantization | Tokens/sec | VRAM used |
|---|---|---|---|
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | 79.28 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.2-1B-Instruct | Q4 | 79.11 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/Llama-3.2-3B-Instruct | Q4 | 79.11 tok/sEstimated Auto-generated benchmark | 2GB |
| inference-net/Schematron-3B | Q4 | 78.40 tok/sEstimated Auto-generated benchmark | 2GB |
| facebook/sam3 | Q4 | 77.88 tok/sEstimated Auto-generated benchmark | 1GB |
| LiquidAI/LFM2-1.2B | Q4 | 77.82 tok/sEstimated Auto-generated benchmark | 1GB |
| apple/OpenELM-1_1B-Instruct | Q4 | 76.93 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-Guard-3-1B | Q4 | 76.44 tok/sEstimated Auto-generated benchmark | 1GB |
| google/gemma-3-1b-it | Q4 | 75.07 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | 74.96 tok/sEstimated Auto-generated benchmark | 2GB |
| bigcode/starcoder2-3b | Q4 | 74.86 tok/sEstimated Auto-generated benchmark | 2GB |
| nari-labs/Dia2-2B | Q4 | 74.40 tok/sEstimated Auto-generated benchmark | 2GB |
| deepseek-ai/DeepSeek-OCR | Q4 | 73.36 tok/sEstimated Auto-generated benchmark | 2GB |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | 73.23 tok/sEstimated Auto-generated benchmark | 1GB |
| unsloth/gemma-3-1b-it | Q4 | 72.18 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-1B | Q4 | 72.08 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-granite/granite-3.3-2b-instruct | Q4 | 71.50 tok/sEstimated Auto-generated benchmark | 1GB |
| ibm-research/PowerMoE-3b | Q4 | 71.44 tok/sEstimated Auto-generated benchmark | 2GB |
| deepseek-ai/deepseek-coder-1.3b-instruct | Q4 | 70.31 tok/sEstimated Auto-generated benchmark | 2GB |
| google/gemma-2-2b-it | Q4 | 69.97 tok/sEstimated Auto-generated benchmark | 1GB |
| tencent/HunyuanOCR | Q4 | 69.93 tok/sEstimated Auto-generated benchmark | 1GB |
| Qwen/Qwen2.5-3B | Q4 | 68.56 tok/sEstimated Auto-generated benchmark | 2GB |
| google/embeddinggemma-300m | Q4 | 67.97 tok/sEstimated Auto-generated benchmark | 1GB |
| google-bert/bert-base-uncased | Q4 | 67.02 tok/sEstimated Auto-generated benchmark | 1GB |
| WeiboAI/VibeThinker-1.5B | Q4 | 66.97 tok/sEstimated Auto-generated benchmark | 1GB |
| meta-llama/Llama-3.2-3B | Q4 | 66.71 tok/sEstimated Auto-generated benchmark | 2GB |
| meta-llama/Llama-3.1-8B-Instruct | Q4 | 66.51 tok/sEstimated Auto-generated benchmark | 4GB |
| google-t5/t5-3b | Q4 | 66.50 tok/sEstimated Auto-generated benchmark | 2GB |
| Qwen/Qwen2.5-1.5B | Q4 | 66.47 tok/sEstimated Auto-generated benchmark | 3GB |
| deepseek-ai/DeepSeek-V3-0324 | Q4 | 66.45 tok/sEstimated Auto-generated benchmark | 4GB |
| meta-llama/Llama-2-7b-hf | Q4 | 66.17 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2.5-3B-Instruct | Q4 | 66.15 tok/sEstimated Auto-generated benchmark | 2GB |
| unsloth/Meta-Llama-3.1-8B-Instruct | Q4 | 66.13 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/DialoGPT-small | Q4 | 66.08 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-4B-Instruct-2507 | Q4 | 66.07 tok/sEstimated Auto-generated benchmark | 2GB |
| zai-org/GLM-4.5-Air | Q4 | 66.03 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/phi-4 | Q4 | 66.00 tok/sEstimated Auto-generated benchmark | 4GB |
| google/gemma-2b | Q4 | 65.99 tok/sEstimated Auto-generated benchmark | 1GB |
| allenai/OLMo-2-0425-1B | Q4 | 65.96 tok/sEstimated Auto-generated benchmark | 1GB |
| lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit | Q4 | 65.95 tok/sEstimated Auto-generated benchmark | 2GB |
| ibm-granite/granite-3.3-8b-instruct | Q4 | 65.86 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen3-1.7B | Q4 | 65.85 tok/sEstimated Auto-generated benchmark | 4GB |
| HuggingFaceTB/SmolLM2-135M | Q4 | 65.76 tok/sEstimated Auto-generated benchmark | 4GB |
| microsoft/VibeVoice-1.5B | Q4 | 65.72 tok/sEstimated Auto-generated benchmark | 3GB |
| swiss-ai/Apertus-8B-Instruct-2509 | Q4 | 65.70 tok/sEstimated Auto-generated benchmark | 4GB |
| Qwen/Qwen2-0.5B-Instruct | Q4 | 65.67 tok/sEstimated Auto-generated benchmark | 3GB |
| Qwen/Qwen3-8B | Q4 | 65.59 tok/sEstimated Auto-generated benchmark | 4GB |
| unsloth/Llama-3.2-1B-Instruct | Q4 | 65.57 tok/sEstimated Auto-generated benchmark | 1GB |
| deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct | Q4 | 65.57 tok/sEstimated Auto-generated benchmark | 4GB |
| parler-tts/parler-tts-large-v1 | Q4 | 65.52 tok/sEstimated Auto-generated benchmark | 4GB |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
| Model | Quantization | Verdict | Estimated speed | VRAM needed |
|---|---|---|---|---|
| meta-llama/Llama-3.2-1B-Instruct | FP16 | Fits comfortably | 27.03 tok/sEstimated | 2GB (have 24GB) |
| MiniMaxAI/MiniMax-M2 | Q4 | Fits comfortably | 63.30 tok/sEstimated | 4GB (have 24GB) |
| MiniMaxAI/MiniMax-M2 | FP16 | Fits comfortably | 25.30 tok/sEstimated | 15GB (have 24GB) |
| Qwen/Qwen2-0.5B | Q4 | Fits comfortably | 62.57 tok/sEstimated | 3GB (have 24GB) |
| google/gemma-2-2b-it | Q4 | Fits comfortably | 69.97 tok/sEstimated | 1GB (have 24GB) |
| codellama/CodeLlama-34b-hf | Q8 | Not supported | 16.21 tok/sEstimated | 35GB (have 24GB) |
| codellama/CodeLlama-34b-hf | FP16 | Not supported | 7.57 tok/sEstimated | 70GB (have 24GB) |
| ibm-research/PowerMoE-3b | FP16 | Fits comfortably | 25.39 tok/sEstimated | 6GB (have 24GB) |
| ibm-granite/granite-3.3-2b-instruct | Q4 | Fits comfortably | 71.50 tok/sEstimated | 1GB (have 24GB) |
| ibm-granite/granite-3.3-2b-instruct | Q8 | Fits comfortably | 51.69 tok/sEstimated | 2GB (have 24GB) |
| WeiboAI/VibeThinker-1.5B | FP16 | Fits comfortably | 25.64 tok/sEstimated | 4GB (have 24GB) |
| meta-llama/Llama-Guard-3-1B | Q4 | Fits comfortably | 76.44 tok/sEstimated | 1GB (have 24GB) |
| meta-llama/Llama-Guard-3-1B | Q8 | Fits comfortably | 52.52 tok/sEstimated | 1GB (have 24GB) |
| openai-community/gpt2 | Q4 | Fits comfortably | 54.59 tok/sEstimated | 4GB (have 24GB) |
| openai-community/gpt2 | Q8 | Fits comfortably | 45.30 tok/sEstimated | 7GB (have 24GB) |
| openai-community/gpt2 | FP16 | Fits comfortably | 24.05 tok/sEstimated | 15GB (have 24GB) |
| Qwen/Qwen2.5-7B-Instruct | Q4 | Fits comfortably | 59.25 tok/sEstimated | 4GB (have 24GB) |
| Qwen/Qwen2.5-7B-Instruct | Q8 | Fits comfortably | 42.94 tok/sEstimated | 7GB (have 24GB) |
| Qwen/Qwen2.5-7B-Instruct | FP16 | Fits comfortably | 22.26 tok/sEstimated | 15GB (have 24GB) |
| Qwen/Qwen3-0.6B | FP16 | Fits comfortably | 21.74 tok/sEstimated | 13GB (have 24GB) |
| meta-llama/Llama-3.1-8B-Instruct | Q8 | Fits comfortably | 40.98 tok/sEstimated | 9GB (have 24GB) |
| Qwen/Qwen3-Embedding-0.6B | Q8 | Fits comfortably | 45.65 tok/sEstimated | 6GB (have 24GB) |
| Qwen/Qwen3-Embedding-0.6B | FP16 | Fits comfortably | 22.17 tok/sEstimated | 13GB (have 24GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q4 | Fits comfortably | 54.90 tok/sEstimated | 3GB (have 24GB) |
| Qwen/Qwen2.5-1.5B-Instruct | Q8 | Fits comfortably | 39.11 tok/sEstimated | 5GB (have 24GB) |
| TinyLlama/TinyLlama-1.1B-Chat-v1.0 | Q4 | Fits comfortably | 73.23 tok/sEstimated | 1GB (have 24GB) |
| meta-llama/Llama-3.2-1B-Instruct | Q8 | Fits comfortably | 46.68 tok/sEstimated | 1GB (have 24GB) |
| Qwen/Qwen2.5-3B-Instruct | FP16 | Fits comfortably | 28.50 tok/sEstimated | 6GB (have 24GB) |
| bigscience/bloomz-560m | Q8 | Fits comfortably | 40.58 tok/sEstimated | 7GB (have 24GB) |
| context-labs/meta-llama-Llama-3.2-3B-Instruct-FP16 | Q4 | Fits comfortably | 79.28 tok/sEstimated | 2GB (have 24GB) |
| RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic | Q8 | Not supported | 16.14 tok/sEstimated | 68GB (have 24GB) |
| RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic | FP16 | Not supported | 7.61 tok/sEstimated | 137GB (have 24GB) |
| meta-llama/Llama-3.2-3B-Instruct | Q4 | Fits comfortably | 74.96 tok/sEstimated | 2GB (have 24GB) |
| meta-llama/Llama-3.2-3B-Instruct | Q8 | Fits comfortably | 51.37 tok/sEstimated | 3GB (have 24GB) |
| Qwen/Qwen2.5-7B | FP16 | Fits comfortably | 21.75 tok/sEstimated | 15GB (have 24GB) |
| Qwen/Qwen2.5-0.5B-Instruct | Q4 | Fits comfortably | 57.76 tok/sEstimated | 3GB (have 24GB) |
| Qwen/Qwen2.5-0.5B-Instruct | Q8 | Fits comfortably | 41.10 tok/sEstimated | 5GB (have 24GB) |
| Qwen/Qwen2.5-0.5B-Instruct | FP16 | Fits comfortably | 21.13 tok/sEstimated | 11GB (have 24GB) |
| google-t5/t5-3b | Q8 | Fits comfortably | 49.02 tok/sEstimated | 3GB (have 24GB) |
| Qwen/Qwen3-Reranker-0.6B | Q8 | Fits comfortably | 43.12 tok/sEstimated | 6GB (have 24GB) |
| Qwen/Qwen3-Reranker-0.6B | FP16 | Fits comfortably | 21.06 tok/sEstimated | 13GB (have 24GB) |
| kaitchup/Phi-3-mini-4k-instruct-gptq-4bit | Q4 | Fits comfortably | 56.67 tok/sEstimated | 2GB (have 24GB) |
| kaitchup/Phi-3-mini-4k-instruct-gptq-4bit | Q8 | Fits comfortably | 41.16 tok/sEstimated | 4GB (have 24GB) |
| Qwen/Qwen2.5-1.5B | Q8 | Fits comfortably | 45.19 tok/sEstimated | 5GB (have 24GB) |
| Qwen/Qwen2.5-1.5B | FP16 | Fits comfortably | 22.97 tok/sEstimated | 11GB (have 24GB) |
| Qwen/Qwen2.5-14B-Instruct | Q4 | Fits comfortably | 41.66 tok/sEstimated | 7GB (have 24GB) |
| Qwen/Qwen2.5-14B-Instruct | Q8 | Fits comfortably | 33.46 tok/sEstimated | 14GB (have 24GB) |
| Qwen/Qwen2.5-0.5B | Q4 | Fits comfortably | 58.36 tok/sEstimated | 3GB (have 24GB) |
| Qwen/Qwen2.5-0.5B | Q8 | Fits comfortably | 46.24 tok/sEstimated | 5GB (have 24GB) |
| meta-llama/Llama-3.1-8B-Instruct | FP16 | Fits comfortably | 25.23 tok/sEstimated | 17GB (have 24GB) |
Note: Performance estimates are calculated. Real results may vary. Methodology · Submit real data
Explore how RTX 5070 stacks up for local inference workloads.
Explore how RTX 4060 Ti 16GB stacks up for local inference workloads.
Explore how RX 6800 XT stacks up for local inference workloads.
Explore how RTX 4070 Super stacks up for local inference workloads.
Explore how RTX 3080 stacks up for local inference workloads.