❌Won't fitfp16 · 4K ctx · batch 1
Does L40S run DeepSeek-V3 671B? No at fp16.
At fp16 precision, DeepSeek-V3 671B (671 B MoE parameters) needs roughly 1428.9 GB of VRAM. L40S only has 48 GB, short by 1380.9 GB. INT4 quantization brings it down to 359.2 GB.
>999%
VRAM
Won't fit
Estimated VRAM (fp16)
1428.9 GB
of 48 GB on L40S· 1380.9 GB short
Tok/sec
17
TFTT
5465ms
Power
350W
Rent on GPUniq marketplace
L40S
Live prices from verified providers · billed hourly, no commitment.
Memory breakdown (fp16)
Shared Weights
74.00 GBExpert Weights
1342.00 GBKV Cache
1.67 GBActivations
6.67 GBFramework Overhead
4.54 GBQuantization comparison
Lower precision = less VRAM with a small quality trade-off. Quality order: FP16 > INT8 > INT4.
| Precision | VRAM | Utilisation | Fits on L40S? |
|---|---|---|---|
| FP16 (full precision) | 1428.9 GB | 2977% | ❌Won't fit |
| INT8 (8-bit) | 715.8 GB | 1491% | ❌Won't fit |
| INT4 / Q4 (4-bit) | 359.2 GB | 748% | ❌Won't fit |
Frequently asked
- Can L40S run DeepSeek-V3 671B?
- Not at fp16 — DeepSeek-V3 671B needs about 1428.9 GB while L40S has 48 GB. It fits at INT4 quantization (359.2 GB) with some accuracy trade-off.
- How much VRAM does DeepSeek-V3 671B use?
- About 1428.9 GB at fp16, 359.2 GB at INT4 (for a 4K context, batch size 1). Longer contexts add to KV cache size; larger batches increase activations.
- What's the fastest way to run DeepSeek-V3 671B on L40S?
- Use a production inference engine like vLLM, SGLang, or TensorRT-LLM with paged attention — they cut KV cache 2–3× vs the naive estimate and batch multiple requests efficiently. For L40S, enable Flash Attention 2.
- Where can I rent a L40S?
- GPUniq aggregates live L40S offers from verified providers. You can deploy an instance in about a minute and pay hourly with no commitment.
Other GPUs for DeepSeek-V3 671B
Try it live
Chat with DeepSeek-V3 671B — no setup
Send a prompt and see how DeepSeek-V3 671B responds directly in our chat. No installation, no GPU required to test.
Want to tweak sequence length, batch size, or fine-tuning? Open the full calculator →
None of the common quantization levels fit this model on L40S. Consider multi-GPU deployment or a larger card.