L40S: 48 GB, what it runs and what it costs
Specs checked · Rental prices from 18m ago · Aliteq Team · How we compute this
The short answer
The NVIDIA L40S has 48 GB of VRAM and 864 GB/s of memory bandwidth (2023). At Q4_K_M quality or better with an 8K context it runs 35 of the open models we track; the largest that fits with room to spare is DeepSeek R1 Distill 70B. The cheapest rental right now is $0.200/hr (Vast.ai, interruptible).
- VRAM
- 48 GB
- Bandwidth
- 864 GB/s
- Board power
- 350 W
- Launched
- 2023
Source: NVIDIA specification · bandwidth not yet confirmed by a second source, so we don't estimate speed from it
Buy it or rent it
| Provider | Type | From, per hour | Rent |
|---|---|---|---|
| Vast.ai | Spot (can be interrupted) | $0.200/hr | Vast.ai |
| Lium | On demand | $0.380/hr | Lium |
| Vast.ai | On demand | $0.467/hr | Vast.ai |
| Lium | Spot (can be interrupted) | $0.630/hr | Lium |
| Runpod | On demand | $1.090/hr | Runpod |
Lowest price per provider in our hourly snapshot (18m ago). Rent buttons are referral links. Every card side by side: cloud GPU prices.
Which AI models run on the L40S
Computed by our VRAM engine for an 8K context: the best quality each model fits at in 48 GB. Speeds are upper-bound estimates from memory bandwidth, not benchmarks.
| Model | Best fit | VRAM used | ~tok/s |
|---|---|---|---|
| Qwen3-Next 80B-A3B (MoE) | Q4_K_M | 46 GB96% · tight | — |
| Qwen3-Coder-Next 80B-A3B (MoE) | Q4_K_M | 46 GB96% · tight | — |
| DeepSeek R1 Distill 70B | Q4_K_M | 43 GB90% | — |
| Mixtral 8x7B (MoE) | Q8_0 | 48 GB100% · tight | — |
| Qwen3.6 35B-A3B (MoE) | Q8_0 | 37 GB76% | — |
| Ornith-1.5 35B-A3B (MoE) | Q8_0 | 37 GB76% | — |
| Qwen3.5 35B-A3B (MoE) | Q8_0 | 37 GB76% | — |
| DeepSeek R1 Distill 32B | Q8_0 | 35 GB73% | — |
| Qwen2.5 Coder 32B | Q8_0 | 35 GB73% | — |
| Qwen2.5 32B Instruct | Q8_0 | 35 GB73% | — |
| Qwen3 32B | Q8_0 | 35 GB73% | — |
| NVIDIA Nemotron 3 Nano 30B-A3B (MoE) | Q8_0 | 32 GB67% | — |
Fits only below Q4 quality: gpt-oss 120B (MoE) (Q2_K), Llama 4 Scout 17B-16E (MoE) (Q2_K), GLM-4.5-Air (MoE) (Q2_K).
Too big for 48 GB (rent a bigger card instead): what it costs to run NVIDIA Nemotron 3 Super 120B-A12B (MoE), what it costs to run Qwen3.5 122B-A10B (MoE), what it costs to run Qwen3.8-Flash-Next (MoE), what it costs to run MiniMax-M2.5 (MoE).
Check a model and context of your own in the VRAM calculator.
AI video on the L40S
Alibaba's Wan 2.2 is an open video model you can run yourself in ComfyUI. With 48 GB the L40S fits:
- Wan 2.2 14B in fp16, or in fp8 at 720p (48 GB and up, per Runpod's Wan 2.2 guide)
- Wan 2.2 14B at 640x640 with fp8 weights (24 GB and up, per Runpod's Wan 2.2 guide)
- Wan 2.2 5B, with ComfyUI's native offloading (8 GB and up, per ComfyUI's Wan 2.2 docs)
What each setup needs, the cheapest card to rent and what one clip costs: cloud GPUs for ComfyUI with Flux and Wan 2.2. Against an app or an API: AI video cost per clip, app vs API vs rented GPU. Rather not run a model yourself? Hosted video models such as Seedance, Veo and Hailuo are compared in Kling AI alternatives, priced per clip.
Cards to compare it with
Every card by VRAM tier: all GPUs for gaming and local AI.