Best AI Models for NVIDIA L40S (48.0GB)
With 48 GB of memory, this is a high-end configuration for local AI. You can comfortably run most open-source LLMs including large 70B parameter models at good quantization levels, making it one of the best setups for serious local AI work.
At this memory tier, nearly every popular open-source model is within reach. You can run Llama 3 70B at Q4_K_M or even Q5_K_M quantization with room to spare, handle coding assistants like DeepSeek Coder 33B at high quality, and easily run any 7B–30B model at full or near-full precision. Context windows remain generous even with larger models, so multi-turn conversations and long-document processing work smoothly.
Runs Well
- 70B models (Llama 3 70B, Qwen 72B) at Q4–Q5
- 30B models at Q6–Q8 quality
- 7B–14B models at full FP16 precision
- Vision models (LLaVA, CogVLM) without compromise
Challenging
- Mixture-of-experts models like Mixtral 8x22B at higher quants
- 120B+ models still require lower quantizations
What LLMs Can NVIDIA L40S Run?
113 models · 5 good
Showing compatibility for NVIDIA L40S
| Model | Quant | VRAM | Speed | Context | Status | Grade |
|---|---|---|---|---|---|---|
Q4_K_M·26.2 t/s tok/s·4K ctx·FAIR FIT | Q4_K_M | 21.4 GB | 26.2 t/s | 4K | FAIR FIT | B60 |
Q4_K_M·42.3 t/s tok/s·131K ctx·EASY RUN | Q4_K_M | 13.3 GB | 42.3 t/s | 131K | EASY RUN | C43 |
Q4_K_M·27.4 t/s tok/s·41K ctx·FAIR FIT | Q4_K_M | 20.5 GB | 27.4 t/s | 41K | FAIR FIT | B58 |
Q4_K_M·26.2 t/s tok/s·4K ctx·FAIR FIT | Q4_K_M | 21.4 GB | 26.2 t/s | 4K | FAIR FIT | B60 |
Q4_K_M·30.0 t/s tok/s·262K ctx·FAIR FIT | Q4_K_M | 18.7 GB | 30.0 t/s | 262K | FAIR FIT | B54 |
Q4_K_M·26.2 t/s tok/s·4K ctx·FAIR FIT | Q4_K_M | 21.4 GB | 26.2 t/s | 4K | FAIR FIT | B60 |
IQ2_M·13.3 t/s tok/s·262K ctx·FAIR FIT | IQ2_M | 42.2 GB | 13.3 t/s | 262K | FAIR FIT | B60 |
Q4_K_M·37.1 t/s tok/s·131K ctx·FAIR FIT | Q4_K_M | 15.1 GB | 37.1 t/s | 131K | FAIR FIT | B47 |
Q4_K_M·31.3 t/s tok/s·8K ctx·FAIR FIT | Q4_K_M | 18.0 GB | 31.3 t/s | 8K | FAIR FIT | B52 |
Q4_K_M·27.0 t/s tok/s·16K ctx·FAIR FIT | Q4_K_M | 20.8 GB | 27.0 t/s | 16K | FAIR FIT | B58 |
Q4_K_M·27.6 t/s tok/s·33K ctx·FAIR FIT | Q4_K_M | 20.3 GB | 27.6 t/s | 33K | FAIR FIT | B57 |
Q4_K_M·27.6 t/s tok/s·33K ctx·FAIR FIT | Q4_K_M | 20.3 GB | 27.6 t/s | 33K | FAIR FIT | B57 |
Q4_K_M·27.4 t/s tok/s·33K ctx·FAIR FIT | Q4_K_M | 20.5 GB | 27.4 t/s | 33K | FAIR FIT | B58 |
BF16·36.5 t/s tok/s·4K ctx·FAIR FIT | BF16 | 15.4 GB | 36.5 t/s | 4K | FAIR FIT | B47 |
Q4_K_M·69.9 t/s tok/s·33K ctx·EASY RUN | Q4_K_M | 8.0 GB | 69.9 t/s | 33K | EASY RUN | C34 |
Q4_K_M·37.8 t/s tok/s·33K ctx·FAIR FIT | Q4_K_M | 14.9 GB | 37.8 t/s | 33K | FAIR FIT | B46 |
NVIDIA L40S Specifications
- Brand
- NVIDIA
- Architecture
- Ada Lovelace
- Compute Capability
- 8.9 (CUDA SM version)
- VRAM
- 48.0 GB GDDR6
- Memory Bandwidth
- 864.0 GB/s
- CUDA Cores
- 18,176
- Tensor Cores
- 568
- FP16 Performance
- 366.50 TFLOPS
- TDP
- 350W
- Release Date
- 2023-08-08
Get Started
GPUs to Consider Over NVIDIA L40S
Similar GPUs and upgrades with more VRAM or higher bandwidth for AI
NVIDIA H100 SXM
NVIDIA · Hopper
NVIDIA A100 80GB SXM
NVIDIA · Ampere
NVIDIA H100 PCIe
NVIDIA · Hopper
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
NVIDIA · Blackwell
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
NVIDIA · Blackwell
AMD Instinct MI210
AMD · CDNA 2
Frequently Asked Questions
- Can NVIDIA L40S run Mixtral 8x7B Instruct v0.1?
Yes, the NVIDIA L40S with 48 GB can run Mixtral 8x7B Instruct v0.1, Mixtral 8x7B v0.1, Falcon 40B, and 1334 other models. 11 models run at excellent quality, and 31 at good quality. Check the compatibility table above for the full list with VRAM usage and estimated speed.
- Is NVIDIA L40S good for AI?
The NVIDIA L40S has 48 GB of GDDR6, making it excellent for running local AI models. It supports 42 models at good quality or better. With 864.0 GB/s memory bandwidth, it delivers fast token generation speeds. This is an enthusiast-grade GPU that handles most popular open-source LLMs.
- How many parameters can NVIDIA L40S handle?
With 48 GB, the NVIDIA L40S supports models from 3B to 70B+ parameters depending on quantization level. At Q4_K_M (the recommended sweet spot), you can fit roughly 80B parameters. This means 7B models at high quality (Q6/Q8) or 30B+ models at Q4.
- What quantization should I use on NVIDIA L40S?
For the best balance of quality and speed on the NVIDIA L40S, 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 L40S for AI inference?
With 864.0 GB/s memory bandwidth, the NVIDIA L40S achieves approximately 125 tokens/sec on a 7B model at Q4_K_M — that's very fast, well above conversational speed. A 14B model runs at ~62 tok/s. Token generation speed scales inversely with model size — smaller models are significantly faster.
tok/s = (864 GB/s ÷ model GB) × efficiency
Smaller models = faster inference. Memory bandwidth is the main bottleneck for token generation speed.
Estimated speed on NVIDIA L40S
~20 tok/s~20 tok/s~20 tok/s~22 tok/sReal-world results typically within ±20%. Speed depends on quantization kernel, batch size, and software stack.
- What's the best model for NVIDIA L40S?
The top-rated models for the NVIDIA L40S are Mixtral 8x7B Instruct v0.1, Mixtral 8x7B v0.1, Falcon 40B. 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 L40S need?
The NVIDIA L40S has a TDP of 350 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 750 W PSU or larger. At this power level, a high-airflow case matters: aim for at least two front intake fans and one rear exhaust, with tidy cabling so hot air isn't trapped around the card. LLM inference sustains full GPU load continuously — longer and more consistently than most gaming workloads — so also make sure your CPU cooler can keep up under combined load.