Skip to content

Deepseek MOE 16B Chat VRAM requirements

How much VRAM Deepseek MOE 16B Chat needs
4-bit weights: 9.2 GBWith 4K context: 10.8 GB8-bit, 4K context: 17.8 GB
Parameters
16.4B
Active per token
2.8B
Longest context
4K
Released
Jan 2024
License
Deepseek
Strongest at
Not rated

Its licence (Deepseek) is not a standard open-source one: check the terms before commercial use.

How much VRAM Deepseek MOE 16B Chat needs

Memory Deepseek MOE 16B Chat needs at each precision and context length
PrecisionWeightsWith 4K
16-bit30.5 GB32.1 GB
8-bit16.2 GB17.8 GB
6-bit12.5 GB14.1 GB
5-bit10.8 GB12.4 GB
4-bit9.2 GB10.8 GB
3-bit7.5 GB9.0 GB
2-bit5.7 GB7.3 GB
1.6-bit3.1 GB4.7 GB

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

GPUs and Macs that run Deepseek MOE 16B Chat

The best way Deepseek MOE 16B Chat runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Deepseek MOE 16B Chat on RTX 4060 report
RTX 5060Too bigPDF: Deepseek MOE 16B Chat on RTX 5060 report
RTX 3070Too bigPDF: Deepseek MOE 16B Chat on RTX 3070 report
RTX 3080 10 GB3-bit4K234 tok/s0.6 sPDF: Deepseek MOE 16B Chat on RTX 3080 10 GB report
RTX 3060 12 GB4-bit4K118 tok/s1.4 sPDF: Deepseek MOE 16B Chat on RTX 3060 12 GB report
Arc B5804-bit4K97 tok/s1.0 sPDF: Deepseek MOE 16B Chat on Arc B580 report
RTX 40704-bit4K153 tok/s0.6 sPDF: Deepseek MOE 16B Chat on RTX 4070 report
RTX 4070 Super4-bit4K153 tok/s0.5 sPDF: Deepseek MOE 16B Chat on RTX 4070 Super report
RTX 50704-bit4K188 tok/s0.6 sPDF: Deepseek MOE 16B Chat on RTX 5070 report
RTX 4060 Ti 16 GB6-bit4K76 tok/s0.8 sPDF: Deepseek MOE 16B Chat on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit4K110 tok/s0.8 sPDF: Deepseek MOE 16B Chat on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit4K90 tok/s0.9 sPDF: Deepseek MOE 16B Chat on Arc A770 16 GB report
Radeon RX 9070 XT6-bit4K117 tok/s0.5 sPDF: Deepseek MOE 16B Chat on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit4K151 tok/s0.4 sPDF: Deepseek MOE 16B Chat on RTX 4070 Ti Super report
RTX 4080 Super6-bit4K162 tok/s0.3 sPDF: Deepseek MOE 16B Chat on RTX 4080 Super report
RTX 5070 Ti6-bit4K186 tok/s0.4 sPDF: Deepseek MOE 16B Chat on RTX 5070 Ti report
RTX 50806-bit4K195 tok/s0.3 sPDF: Deepseek MOE 16B Chat on RTX 5080 report
Radeon RX 7900 XTX8-bit4K129 tok/s0.8 sPDF: Deepseek MOE 16B Chat on Radeon RX 7900 XTX report
RTX 30908-bit4K160 tok/s0.5 sPDF: Deepseek MOE 16B Chat on RTX 3090 report
RTX 40908-bit4K168 tok/s0.2 sPDF: Deepseek MOE 16B Chat on RTX 4090 report
Radeon AI PRO R97008-bit4K97 tok/s0.5 sPDF: Deepseek MOE 16B Chat on Radeon AI PRO R9700 report
RTX 50908-bit4K245 tok/s0.2 sPDF: Deepseek MOE 16B Chat on RTX 5090 report
2× RTX 30908-bit4K160 tok/s0.3 sPDF: Deepseek MOE 16B Chat on 2× RTX 3090 report
2× RTX 40908-bit4K168 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K205 tok/s0.3 sPDF: Deepseek MOE 16B Chat on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K163 tok/s0.2 sPDF: Deepseek MOE 16B Chat on RTX 6000 Ada report
L40S8-bit4K150 tok/s0.3 sPDF: Deepseek MOE 16B Chat on L40S report
2× RTX 50908-bit4K245 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 2× RTX 5090 report
4× RTX 30908-bit4K160 tok/s0.2 sPDF: Deepseek MOE 16B Chat on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K245 tok/s0.1 sPDF: Deepseek MOE 16B Chat on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB4-bit4K47 tok/s12 sPDF: Deepseek MOE 16B Chat on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit4K36 tok/s12 sPDF: Deepseek MOE 16B Chat on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K29 tok/s12 sPDF: Deepseek MOE 16B Chat on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit4K70 tok/s5.8 sPDF: Deepseek MOE 16B Chat on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K58 tok/s5.8 sPDF: Deepseek MOE 16B Chat on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K58 tok/s5.8 sPDF: Deepseek MOE 16B Chat on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K89 tok/s2.9 sPDF: Deepseek MOE 16B Chat on M4 Max 128 GB report
M5 Max 128 GB8-bit4K96 tok/s0.8 sPDF: Deepseek MOE 16B Chat on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K105 tok/s1.6 sPDF: Deepseek MOE 16B Chat on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K105 tok/s1.6 sPDF: Deepseek MOE 16B Chat on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K105 tok/s1.6 sPDF: Deepseek MOE 16B Chat on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K129 tok/s0.5 sPDF: Deepseek MOE 16B Chat on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K45 tok/s1.7 sPDF: Deepseek MOE 16B Chat on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K45 tok/s1.7 sPDF: Deepseek MOE 16B Chat on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K63 tok/s0.6 sPDF: Deepseek MOE 16B Chat on RTX Spark report
DGX Spark8-bit4K57 tok/s0.6 sPDF: Deepseek MOE 16B Chat on DGX Spark report
Dual DGX Spark8-bit4K57 tok/s0.3 sPDF: Deepseek MOE 16B Chat on Dual DGX Spark report
DGX Station GB3008-bit4K438 tok/s0.1 sPDF: Deepseek MOE 16B Chat on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K245 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K245 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K245 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K264 tok/s0.4 sPDF: Deepseek MOE 16B Chat on A100 80 GB SXM report
H100 SXM8-bit4K338 tok/s0.1 sPDF: Deepseek MOE 16B Chat on H100 SXM report
4× H200 NVL8-bit4K388 tok/s0.1 sPDF: Deepseek MOE 16B Chat on 4× H200 NVL report
8× H200 NVL8-bit4K388 tok/s0.1 sPDF: Deepseek MOE 16B Chat 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 mradermacher/deepseek-moe-16b-chat-i1-GGUF
vLLM
vllm serve deepseek-ai/deepseek-moe-16b-chat

Questions

How much VRAM does Deepseek MOE 16B Chat need?

Deepseek MOE 16B Chat needs 9.2 GB for its weights at 4-bit, 10.8 GB with an 4K context and 10.8 GB with 4K. At 8-bit, 17.8 GB.

What is the minimum VRAM for Deepseek MOE 16B Chat?

About 9.0 GB at 3-bit with an 4K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 3080 10 GB.

How big is Deepseek MOE 16B Chat?

16.4B parameters (2.8B active per token): 30.5 GB of weights at 16-bit, 9.2 GB at 4-bit.

Can Deepseek MOE 16B Chat run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 168 tokens a second.

More models