Skip to content

Best GPU for DeepSeek V3: VRAM Requirements

DeepSeek V3 needs about 1,476.8 GB of GPU memory in BF16 and 369.7 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.

DeepSeek V3 needs about 1,476.8 GB of VRAM in FP16, 738.7 GB in FP8 and 369.7 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on 8 × NVIDIA B300 288GB; a 4-bit build fits on 2 × NVIDIA B300 288GB. Its weights are published in FP8, so the FP16 figure is for reference.

DeepSeek V3 is a mixture-of-experts language model with 671B total and 37B active parameters per token from DeepSeek, released in December 2024, with a 131,072-token context window and published under the MIT (V3 0324 checkpoint). Its weights alone take 1,342 GB in BF16 and 335.5 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.

Recommended GPUs for DeepSeek V3

DeepSeek V3 VRAM requirements by precision

One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.

PrecisionWeightsKV cacheTotal for inference
BF16 / FP161,342 GB0.6 GB1,476.8 GB
FP8 (published weights)671 GB0.6 GB738.7 GB
INT4335.5 GB0.6 GB369.7 GB

DeepSeek V3 VRAM by task

Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.

TaskGPU memory
Inference at the full 131,072-token context (FP8)748.2 GB
QLoRA (4-bit base)436.2 GB
LoRA (BF16 base)1,744.6 GB
Full fine-tuning (BF16, AdamW)11,809.6 GB

GPUs that fit DeepSeek V3

Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.

GPUMemoryBF16 inference (1,476.8 GB)FP8 inference (738.7 GB)INT4 inference (369.7 GB)QLoRA (436.2 GB)LoRA (1,744.6 GB)
NVIDIA L40S 48GB48 GB——8 GPUs——
NVIDIA Quadro RTX 8000 48GB48 GB——8 GPUs——
NVIDIA RTX 6000 Ada 48GB48 GB——8 GPUs——
NVIDIA RTX A6000 48GB48 GB——8 GPUs——
NVIDIA A100 PCIe 80GB80 GB——8 GPUs8 GPUs—
NVIDIA A100 SXM 80GB80 GB——8 GPUs8 GPUs—
NVIDIA H100 PCIe 80GB80 GB——8 GPUs8 GPUs—
NVIDIA H100 SXM 80GB80 GB——8 GPUs8 GPUs—
NVIDIA RTX PRO 6000 Blackwell 96GB96 GB—8 GPUs4 GPUs8 GPUs—
NVIDIA H200 141GB141 GB—8 GPUs4 GPUs4 GPUs—
NVIDIA B200 180GB180 GB—8 GPUs4 GPUs4 GPUs—
NVIDIA B300 288GB288 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA GB300 NVL288 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs

What is DeepSeek V3?

DeepSeek V3 is DeepSeek's general chat model, a 671-billion-parameter mixture-of-experts model with 37 billion active parameters per token and multi-head latent attention. Its March 2025 checkpoint, DeepSeek V3 0324, improves reasoning and front-end coding and is published in FP8 under the MIT license.

Developer
DeepSeek
Released
December 2024
Parameters
671B total, 37B active per token
Context window
131,072 tokens
License
MIT (V3 0324 checkpoint)

Multi-head latent attention caches one compressed 576-value vector per layer and token, counted as a KV width of 288. The checkpoint also holds a 14-billion-parameter multi-token prediction module, not counted here.

Source: DeepSeek V3 0324 model card (as of November 17, 2026).

Frequently asked questions

How much VRAM does DeepSeek V3 need?
For inference with an 8,192-token context and one request, DeepSeek V3 needs about 1,476.8 GB in BF16, 738.7 GB in FP8, 369.7 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
Can DeepSeek V3 run on a single GPU?
Not among the GPUs listed here: even a 4-bit build needs about 369.7 GB, so the model is split across several GPUs with tensor or pipeline parallelism.
How much GPU memory does fine-tuning DeepSeek V3 take?
QLoRA (4-bit base): about 436.2 GB; LoRA (BF16 base): about 1,744.6 GB; Full fine-tuning (BF16, AdamW): about 11,809.6 GB. Longer training sequences and larger batches add activation memory on top.
What is the best GPU for DeepSeek V3?
Best for inference: 4 GPUs × NVIDIA B300 288GB (FP8 or INT8 weights, 738.7 GB at 8,192 tokens). Best for fine-tuning: 8 GPUs × NVIDIA B300 288GB (LoRA on BF16 weights, 1,744.6 GB). Budget pick: 8 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB (Smallest card class that holds a 4-bit build, 369.7 GB). Check live offers before booking, as availability changes.
Can DeepSeek V3 run on an RTX 4090?
No: even a 4-bit build needs about 369.7 GB, more than eight RTX 4090 cards hold (192 GB). It needs data center GPUs with more memory per card.
Can DeepSeek V3 run on an RTX 5090?
No: even a 4-bit build needs about 369.7 GB, more than eight RTX 5090 cards hold (256 GB). It needs data center GPUs with more memory per card.

Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.

Explore the live market

Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.