Skip to content

DeepSeek V3.2 VRAM requirements

How much VRAM DeepSeek V3.2 needs
4-bit weights: 379 GBWith 8K context: 380 GBWith 32K context: 382 GB
Parameters
671B
Active per token
37B
Longest context
160K
Released
Dec 2025
License
MIT
Strongest at
Chat, Coding, Agents, Docs

How much VRAM DeepSeek V3.2 needs

Memory DeepSeek V3.2 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
As released (FP8)642 GB644 GB645 GB653 GB
6-bit512 GB514 GB516 GB523 GB
5-bit444 GB446 GB448 GB455 GB
4-bit379 GB380 GB382 GB389 GB
3-bit305 GB307 GB309 GB316 GB
2-bit234 GB236 GB238 GB245 GB
1.6-bit129 GB130 GB132 GB139 GB

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

GPUs and Macs that run DeepSeek V3.2

The best way DeepSeek V3.2 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: DeepSeek V3.2 on RTX 4060 report
RTX 5060Too bigPDF: DeepSeek V3.2 on RTX 5060 report
RTX 3070Too bigPDF: DeepSeek V3.2 on RTX 3070 report
RTX 3080 10 GBToo bigPDF: DeepSeek V3.2 on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: DeepSeek V3.2 on RTX 3060 12 GB report
Arc B580Too bigPDF: DeepSeek V3.2 on Arc B580 report
RTX 4070Too bigPDF: DeepSeek V3.2 on RTX 4070 report
RTX 4070 SuperToo bigPDF: DeepSeek V3.2 on RTX 4070 Super report
RTX 5070Too bigPDF: DeepSeek V3.2 on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: DeepSeek V3.2 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: DeepSeek V3.2 on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: DeepSeek V3.2 on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: DeepSeek V3.2 on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: DeepSeek V3.2 on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: DeepSeek V3.2 on RTX 4080 Super report
RTX 5070 TiToo bigPDF: DeepSeek V3.2 on RTX 5070 Ti report
RTX 5080Too bigPDF: DeepSeek V3.2 on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: DeepSeek V3.2 on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: DeepSeek V3.2 on RTX 3090 report
RTX 4090Too bigPDF: DeepSeek V3.2 on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: DeepSeek V3.2 on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: DeepSeek V3.2 on RTX 5090 report
2× RTX 3090Too bigPDF: DeepSeek V3.2 on 2× RTX 3090 report
2× RTX 4090Too bigPDF: DeepSeek V3.2 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GBToo bigPDF: DeepSeek V3.2 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 AdaToo bigPDF: DeepSeek V3.2 on RTX 6000 Ada report
L40SToo bigPDF: DeepSeek V3.2 on L40S report
2× RTX 5090Too bigPDF: DeepSeek V3.2 on 2× RTX 5090 report
4× RTX 3090Too bigPDF: DeepSeek V3.2 on 4× RTX 3090 report
RTX PRO 6000 BlackwellToo bigPDF: DeepSeek V3.2 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GBToo bigPDF: DeepSeek V3.2 on Mac mini M4 Pro 64 GB report
M4 Max 128 GBToo bigPDF: DeepSeek V3.2 on M4 Max 128 GB report
M5 Max 128 GBToo bigPDF: DeepSeek V3.2 on M5 Max 128 GB report
M3 Ultra 96 GBToo bigPDF: DeepSeek V3.2 on M3 Ultra 96 GB report
M3 Ultra 256 GB1.6-bit160K44 tok/s23 sPDF: DeepSeek V3.2 on M3 Ultra 256 GB report
M3 Ultra 512 GB3-bit160K24 tok/s23 sPDF: DeepSeek V3.2 on M3 Ultra 512 GB report
M5 Ultra 512 GB3-bit160K32 tok/s7.2 sPDF: DeepSeek V3.2 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GBToo bigPDF: DeepSeek V3.2 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GBToo bigPDF: DeepSeek V3.2 on Ryzen AI Max+ 395 128 GB report
RTX SparkToo bigPDF: DeepSeek V3.2 on RTX Spark report
DGX SparkToo bigPDF: DeepSeek V3.2 on DGX Spark report
Dual DGX Spark1.6-bit160K23 tok/s4.7 sPDF: DeepSeek V3.2 on Dual DGX Spark report
DGX Station GB3002-bit52K155 tok/s0.7 sPDF: DeepSeek V3.2 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell1.6-bit160K103 tok/s1.1 sPDF: DeepSeek V3.2 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell3-bit160K55 tok/s0.6 sPDF: DeepSeek V3.2 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 BlackwellFP8160K29 tok/s0.4 sPDF: DeepSeek V3.2 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXMToo bigPDF: DeepSeek V3.2 on A100 80 GB SXM report
H100 SXMToo bigPDF: DeepSeek V3.2 on H100 SXM report
4× H200 NVL6-bit160K79 tok/s0.6 sPDF: DeepSeek V3.2 on 4× H200 NVL report
8× H200 NVLFP8160K67 tok/s0.3 sPDF: DeepSeek V3.2 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-V3.2-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-V3.2
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-V3.2-4bit

Questions

How much VRAM does DeepSeek V3.2 need?

DeepSeek V3.2 needs 379 GB for its weights at 4-bit, 380 GB with an 8K context and 382 GB with 32K.

What is the minimum VRAM for DeepSeek V3.2?

About 130 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 V3.2?

671B parameters (37B active per token): 642 GB of weights as released (FP8), 379 GB at 4-bit.

Can DeepSeek V3.2 run on an RTX 4090?

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

More models