L40S Price & Used Prices: Which Local LLMs Can It Run?

A single L40S has 48GB of GDDR6 VRAM. At 4-bit quantization and 8K context it fits a dense model of up to about 79B parameters, or about 43B at 8-bit.

Yes. At 4-bit with an 8K context a 70B dense model needs about 43GB, which fits in the L40S's 48GB.

Among models released in 2026, the largest dense model that fits entirely in VRAM at 4-bit is Gemma 4 31B, at an estimated 32.5 tokens/s. Among MoE models, Qwen3.6 35B A3B is the fastest at about 142.2 tokens/s. The full list is in the table above.

Estimated from memory bandwidth at 4-bit and 8K context: 8B dense: about 109.1 tokens/s, 14B dense: about 66 tokens/s, 32B dense: about 30.7 tokens/s, and 70B dense: about 14.5 tokens/s. MoE models read only the active parameters for each token, so they run much faster. Real-world results are usually 70%–100% of the estimate.

With vLLM tensor parallelism at 4-bit and 32K context, two L40S cards cannot fit any model of 50B or larger released in 2026. At 4-bit with 32K context the minimum is 4 cards, which fits Qwen3.8 Flash Next and DeepSeek V4 Flash.

As of Sep 29, 2026: eBay median $8,974 (18 listings). Figures are medians of live listings, refreshed every 6 hours.

As of Sep 29, 2026: RunPod at $0.79 per hour and Vast.ai at $0.8 per hour. All are single-GPU on-demand rates, excluding storage and data transfer.

Taking the eBay median of $8,974 and electricity at $0.18 per kWh, at 8 hours a day buying pays for itself in about 4.2 years (versus RunPod at $0.79 per hour). If you will use it for longer than 4.2 years, buy; otherwise, rent.

The RTX 6000 Ada has 48GB of VRAM and 960 GB/s of bandwidth; the L40S has 48GB and 864 GB/s. For 70B dense models at 4-bit, the L40S is estimated at about 14.5 tokens/s and the RTX 6000 Ada at about 16.1 tokens/s.