Skip to content

DeepSeek V3 VRAM requirements

How much VRAM DeepSeek V3 needs
4-bit weights: 386 GBWith 8K context: 388 GBWith 32K context: 390 GB
Parameters
684B
Active per token
40.1B
Longest context
160K
Released
Dec 2024
Strongest at
Not rated

How much VRAM DeepSeek V3 needs

Memory DeepSeek V3 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
As released (FP8)641 GB643 GB644 GB651 GB
6-bit523 GB524 GB526 GB532 GB
5-bit453 GB455 GB456 GB463 GB
4-bit386 GB388 GB390 GB396 GB
3-bit312 GB313 GB315 GB321 GB
2-bit239 GB241 GB242 GB249 GB
1.6-bit131 GB133 GB135 GB141 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run DeepSeek V3

The best way DeepSeek V3 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: DeepSeek V3 on RTX 4060 report
RTX 5060Too bigPDF: DeepSeek V3 on RTX 5060 report
RTX 3070Too bigPDF: DeepSeek V3 on RTX 3070 report
RTX 3080 10 GBToo bigPDF: DeepSeek V3 on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: DeepSeek V3 on RTX 3060 12 GB report
Arc B580Too bigPDF: DeepSeek V3 on Arc B580 report
RTX 4070Too bigPDF: DeepSeek V3 on RTX 4070 report
RTX 4070 SuperToo bigPDF: DeepSeek V3 on RTX 4070 Super report
RTX 5070Too bigPDF: DeepSeek V3 on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: DeepSeek V3 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: DeepSeek V3 on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: DeepSeek V3 on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: DeepSeek V3 on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: DeepSeek V3 on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: DeepSeek V3 on RTX 4080 Super report
RTX 5070 TiToo bigPDF: DeepSeek V3 on RTX 5070 Ti report
RTX 5080Too bigPDF: DeepSeek V3 on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: DeepSeek V3 on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: DeepSeek V3 on RTX 3090 report
RTX 4090Too bigPDF: DeepSeek V3 on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: DeepSeek V3 on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: DeepSeek V3 on RTX 5090 report
2× RTX 3090Too bigPDF: DeepSeek V3 on 2× RTX 3090 report
2× RTX 4090Too bigPDF: DeepSeek V3 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GBToo bigPDF: DeepSeek V3 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 AdaToo bigPDF: DeepSeek V3 on RTX 6000 Ada report
L40SToo bigPDF: DeepSeek V3 on L40S report
2× RTX 5090Too bigPDF: DeepSeek V3 on 2× RTX 5090 report
4× RTX 3090Too bigPDF: DeepSeek V3 on 4× RTX 3090 report
RTX PRO 6000 BlackwellToo bigPDF: DeepSeek V3 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: DeepSeek V3 on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: DeepSeek V3 on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: DeepSeek V3 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: DeepSeek V3 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: DeepSeek V3 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GBToo bigPDF: DeepSeek V3 on Mac mini M4 Pro 64 GB report
M4 Max 128 GBToo bigPDF: DeepSeek V3 on M4 Max 128 GB report
M5 Max 128 GBToo bigPDF: DeepSeek V3 on M5 Max 128 GB report
M3 Ultra 96 GBToo bigPDF: DeepSeek V3 on M3 Ultra 96 GB report
M3 Ultra 256 GB1.6-bit160K42 tok/s25 sPDF: DeepSeek V3 on M3 Ultra 256 GB report
M3 Ultra 512 GB3-bit160K22 tok/s25 sPDF: DeepSeek V3 on M3 Ultra 512 GB report
M5 Ultra 512 GB3-bit160K30 tok/s7.8 sPDF: DeepSeek V3 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GBToo bigPDF: DeepSeek V3 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GBToo bigPDF: DeepSeek V3 on Ryzen AI Max+ 395 128 GB report
RTX SparkToo bigPDF: DeepSeek V3 on RTX Spark report
DGX SparkToo bigPDF: DeepSeek V3 on DGX Spark report
Dual DGX Spark1.6-bit160K21 tok/s5.1 sPDF: DeepSeek V3 on Dual DGX Spark report
DGX Station GB3002-bit160K150 tok/s0.7 sPDF: DeepSeek V3 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell1.6-bit160K98 tok/s1.2 sPDF: DeepSeek V3 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell3-bit160K52 tok/s0.7 sPDF: DeepSeek V3 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 BlackwellFP8160K28 tok/s0.4 sPDF: DeepSeek V3 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXMToo bigPDF: DeepSeek V3 on A100 80 GB SXM report
H100 SXMToo bigPDF: DeepSeek V3 on H100 SXM report
4× H200 NVL6-bit144K74 tok/s0.7 sPDF: DeepSeek V3 on 4× H200 NVL report
8× H200 NVLFP8160K64 tok/s0.3 sPDF: DeepSeek V3 on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf unsloth/DeepSeek-V3-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-V3
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-V3-4bit

Questions

How much VRAM does DeepSeek V3 need?

DeepSeek V3 needs 386 GB for its weights at 4-bit, 388 GB with an 8K context and 390 GB with 32K.

What is the minimum VRAM for DeepSeek V3?

About 133 GB at 1.6-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the M3 Ultra 256 GB.

How big is DeepSeek V3?

684B parameters (40.1B active per token): 641 GB of weights as released (FP8), 386 GB at 4-bit.

Can DeepSeek V3 run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the M3 Ultra 256 GB.

More models