Skip to content

Deepseek Coder 6.7B Base VRAM requirements

How much VRAM Deepseek Coder 6.7B Base needs
4-bit weights: 3.8 GBWith 8K context: 8.4 GBWith 16K context: 12.4 GB8-bit, 8K context: 11.3 GB
Parameters
6.7B
Longest context
16K
Released
Oct 2023
License
Deepseek License
Strongest at
Coding

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

How much VRAM Deepseek Coder 6.7B Base needs

Memory Deepseek Coder 6.7B Base needs at each precision and context length
PrecisionWeightsWith 8KWith 16K
16-bit12.6 GB17.2 GB21.2 GB
8-bit6.7 GB11.3 GB15.3 GB
6-bit5.1 GB9.8 GB13.8 GB
5-bit4.5 GB9.1 GB13.1 GB
4-bit3.8 GB8.4 GB12.4 GB
3-bit3.1 GB7.7 GB11.7 GB
2-bit2.4 GB7.0 GB11.0 GB
1.6-bit1.3 GB5.9 GB9.9 GB

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

GPUs and Macs that run Deepseek Coder 6.7B Base

The best way Deepseek Coder 6.7B Base runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Deepseek Coder 6.7B Base on RTX 4060 report
RTX 5060Too bigPDF: Deepseek Coder 6.7B Base on RTX 5060 report
RTX 3070Too bigPDF: Deepseek Coder 6.7B Base on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Deepseek Coder 6.7B Base on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Deepseek Coder 6.7B Base on RTX 3060 12 GB report
Arc B580Too bigPDF: Deepseek Coder 6.7B Base on Arc B580 report
RTX 4070Too bigPDF: Deepseek Coder 6.7B Base on RTX 4070 report
RTX 4070 SuperToo bigPDF: Deepseek Coder 6.7B Base on RTX 4070 Super report
RTX 5070Too bigPDF: Deepseek Coder 6.7B Base on RTX 5070 report
RTX 4060 Ti 16 GB6-bit16K36 tok/s0.7 sPDF: Deepseek Coder 6.7B Base on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit16K55 tok/s0.6 sPDF: Deepseek Coder 6.7B Base on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit16K47 tok/s0.7 sPDF: Deepseek Coder 6.7B Base on Arc A770 16 GB report
Radeon RX 9070 XT6-bit16K64 tok/s0.4 sPDF: Deepseek Coder 6.7B Base on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit16K80 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on RTX 4070 Ti Super report
RTX 4080 Super6-bit16K87 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on RTX 4080 Super report
RTX 5070 Ti6-bit16K104 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on RTX 5070 Ti report
RTX 50806-bit16K111 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on RTX 5080 report
Radeon RX 7900 XTX8-bit16K73 tok/s0.6 sPDF: Deepseek Coder 6.7B Base on Radeon RX 7900 XTX report
RTX 30908-bit16K86 tok/s0.4 sPDF: Deepseek Coder 6.7B Base on RTX 3090 report
RTX 40908-bit16K92 tok/s0.2 sPDF: Deepseek Coder 6.7B Base on RTX 4090 report
Radeon AI PRO R97008-bit16K50 tok/s0.4 sPDF: Deepseek Coder 6.7B Base on Radeon AI PRO R9700 report
RTX 50908-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on RTX 5090 report
2× RTX 30908-bit16K86 tok/s0.2 sPDF: Deepseek Coder 6.7B Base on 2× RTX 3090 report
2× RTX 40908-bit16K92 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit16K119 tok/s0.2 sPDF: Deepseek Coder 6.7B Base on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit16K88 tok/s0.2 sPDF: Deepseek Coder 6.7B Base on RTX 6000 Ada report
L40S8-bit16K80 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on L40S report
2× RTX 50908-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 2× RTX 5090 report
4× RTX 30908-bit16K86 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Deepseek Coder 6.7B Base on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit16K14 tok/s9.2 sPDF: Deepseek Coder 6.7B Base on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit16K14 tok/s9.2 sPDF: Deepseek Coder 6.7B Base on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit16K29 tok/s4.6 sPDF: Deepseek Coder 6.7B Base on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit16K29 tok/s4.6 sPDF: Deepseek Coder 6.7B Base on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit16K29 tok/s4.6 sPDF: Deepseek Coder 6.7B Base on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit16K51 tok/s2.3 sPDF: Deepseek Coder 6.7B Base on M4 Max 128 GB report
M5 Max 128 GB8-bit16K56 tok/s0.7 sPDF: Deepseek Coder 6.7B Base on M5 Max 128 GB report
M3 Ultra 96 GB8-bit16K63 tok/s1.3 sPDF: Deepseek Coder 6.7B Base on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit16K63 tok/s1.3 sPDF: Deepseek Coder 6.7B Base on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit16K63 tok/s1.3 sPDF: Deepseek Coder 6.7B Base on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit16K86 tok/s0.4 sPDF: Deepseek Coder 6.7B Base on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit16K21 tok/s1.3 sPDF: Deepseek Coder 6.7B Base on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit16K21 tok/s1.3 sPDF: Deepseek Coder 6.7B Base on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit16K29 tok/s0.5 sPDF: Deepseek Coder 6.7B Base on RTX Spark report
DGX Spark8-bit16K27 tok/s0.5 sPDF: Deepseek Coder 6.7B Base on DGX Spark report
Dual DGX Spark8-bit16K27 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on Dual DGX Spark report
DGX Station GB3008-bit16K421 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit16K152 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit16K170 tok/s0.3 sPDF: Deepseek Coder 6.7B Base on A100 80 GB SXM report
H100 SXM8-bit16K253 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on H100 SXM report
4× H200 NVL8-bit16K328 tok/s0.1 sPDF: Deepseek Coder 6.7B Base on 4× H200 NVL report
8× H200 NVL8-bit16K328 tok/s0.1 sPDF: Deepseek Coder 6.7B Base 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 TheBloke/deepseek-coder-6.7B-base-GGUF
vLLM
vllm serve deepseek-ai/deepseek-coder-6.7b-base

Questions

How much VRAM does Deepseek Coder 6.7B Base need?

Deepseek Coder 6.7B Base needs 3.8 GB for its weights at 4-bit, 8.4 GB with an 8K context and 12.4 GB with 16K. At 8-bit, 11.3 GB.

What is the minimum VRAM for Deepseek Coder 6.7B Base?

About 7.7 GB at 3-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 RTX 4060 Ti 16 GB.

How big is Deepseek Coder 6.7B Base?

6.7B parameters: 12.6 GB of weights at 16-bit, 3.8 GB at 4-bit.

Can Deepseek Coder 6.7B Base run on an RTX 4090?

Yes. At 8-bit it holds 16K of context and writes about 92 tokens a second.

More models