NVIDIA L40S for local LLMs
Written by Jakub Rusinowski · Last updated
With 48 GB at 864 GB/s, the L40S runs 118 catalogued models at Q4_K_M with 8K context. The largest that fits is Llama 3.3 70B Instruct (~45.7 GB), and the top pick is Cosmos 3 Super at 23–45 tok/s.
The mainstream rack-server inference card of the Ada generation: 48 GB GDDR6, passive cooling (needs server chassis airflow), widely available in 1–8× configurations from Dell, HPE, Lenovo, and Supermicro. A common building block for on-prem AI pilots because it slots into standard 2U servers without exotic power or cooling.
Models that run on the L40S
Q4_K_M, 8K context, 48 GB usable. Ranked by quality and speed.
| Model | Memory · marker = 48 GB | VRAM | Speed |
|---|---|---|---|
| Cosmos 3 Super Cosmos 3 | 39.4 GB | 23–45 tok/s | |
| K2 Horizon MoVA 36B-A4B K2 Horizon | 25 GB | 94–181 tok/s | |
| Qwen 3.6 35B-A3B Qwen 3.6 | 22.1 GB | 145–279 tok/s | |
| Nex-N2.5 mini Nex-N2.5 | 21.9 GB | 100–192 tok/s | |
| Nex-N2 mini Nex-N2 | 21.9 GB | 100–192 tok/s | |
| Qwen 3.5 35B-A3B Qwen 3.5 | 21.9 GB | 100–192 tok/s | |
| Laguna XS 2.1 33B-A3B Poolside Laguna XS 2.1 | 20.7 GB | 101–193 tok/s | |
| Qwen 3 32B Qwen 3 | 22.8 GB | 23–44 tok/s | |
| Aya Expanse 32B Aya Expanse | 21.6 GB | 24–47 tok/s | |
| EXAONE 3.5 32B EXAONE 3.5 | 22.3 GB | 24–45 tok/s | |
| Cogito v1 32B Cogito v1 | 22.3 GB | 24–45 tok/s | |
| Granite 4.0 Small-H 32B-A9B IBM Granite 4.0 | 20.1 GB | 60–116 tok/s |
Buy it or rent the same memory
Buy the card, or rent a GPU with the same memory by the hour to try models first.
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
Speed vs other GPUs
Llama 3.1 8B, Q4_K_M. Estimated ranges. How this is calculated
Specifications
Specs last updated 2026-07-21.
- Memory
- 48 GB
- Memory bandwidth
- 864 GB/s
- Architecture
- Ada Lovelace AD102
- Series
- Server
- Board power
- 350 W
- Release year
- 2023
- Launch price
- $7,499
- Usable for models
- 48 GB
Similar GPUs
Frequently asked questions
Can the NVIDIA L40S run local LLMs?
Yes. With 48 GB (48 GB usable by a model) it runs 118 of the catalogued models at Q4_K_M with 8K context; the largest is Llama 3.3 70B Instruct, needing about 45.7 GB.
How fast is the NVIDIA L40S for AI inference?
It is estimated to run Llama 3.1 8B at 77–147 tok/s at Q4_K_M. Llama 3.3 70B is estimated at 12–23 tok/s. These are modelled estimates from memory bandwidth, not measurements; the methodology page shows the formula.
What LLMs can I run on 48 GB?
Among the best that fit: Cosmos 3 Super, K2 Horizon MoVA 36B-A4B, Qwen 3.6 35B-A3B, Nex-N2.5 mini, Nex-N2 mini. The quickest start is Ollama: ollama run llama3.3:70b.
Keep going
Can I run it on the L40S?
More for this card
Tools