Skip to content

DeepSeek Coder V2 Lite Instruct VRAM requirements

How much VRAM DeepSeek Coder V2 Lite Instruct needs
4-bit weights: 8.9 GBWith 8K context: 9.8 GBWith 32K context: 10.5 GB8-bit, 8K context: 16.5 GB
Parameters
15.7B
Active per token
2.7B
Longest context
160K
Released
Jun 2024
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 V2 Lite Instruct needs

Memory DeepSeek Coder V2 Lite Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit29.3 GB30.2 GB30.9 GB33.8 GB
8-bit15.5 GB16.5 GB17.2 GB20.0 GB
6-bit12.0 GB12.9 GB13.7 GB16.5 GB
5-bit10.4 GB11.3 GB12.1 GB14.9 GB
4-bit8.9 GB9.8 GB10.5 GB13.4 GB
3-bit7.1 GB8.1 GB8.8 GB11.7 GB
2-bit5.5 GB6.4 GB7.1 GB10.0 GB
1.6-bit3.0 GB4.0 GB4.7 GB7.5 GB

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

GPUs and Macs that run DeepSeek Coder V2 Lite Instruct

The best way DeepSeek Coder V2 Lite Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: DeepSeek Coder V2 Lite Instruct on RTX 4060 report
RTX 5060Too bigPDF: DeepSeek Coder V2 Lite Instruct on RTX 5060 report
RTX 3070Too bigPDF: DeepSeek Coder V2 Lite Instruct on RTX 3070 report
RTX 3080 10 GB3-bit52K246 tok/s0.6 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 3080 10 GB report
RTX 3060 12 GB4-bit60K125 tok/s1.3 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 3060 12 GB report
Arc B5804-bit60K102 tok/s0.9 sPDF: DeepSeek Coder V2 Lite Instruct on Arc B580 report
RTX 40704-bit60K161 tok/s0.6 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4070 report
RTX 4070 Super4-bit60K161 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4070 Super report
RTX 50704-bit60K198 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 5070 report
RTX 4060 Ti 16 GB6-bit80K80 tok/s0.8 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit80K117 tok/s0.7 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit80K95 tok/s0.8 sPDF: DeepSeek Coder V2 Lite Instruct on Arc A770 16 GB report
Radeon RX 9070 XT6-bit80K123 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit80K160 tok/s0.4 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4070 Ti Super report
RTX 4080 Super6-bit80K171 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4080 Super report
RTX 5070 Ti6-bit80K196 tok/s0.4 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 5070 Ti report
RTX 50806-bit80K205 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit160K136 tok/s0.7 sPDF: DeepSeek Coder V2 Lite Instruct on Radeon RX 7900 XTX report
RTX 30908-bit160K168 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 3090 report
RTX 40908-bit160K178 tok/s0.2 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 4090 report
Radeon AI PRO R97008-bit160K102 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on Radeon AI PRO R9700 report
RTX 50908-bit160K258 tok/s0.2 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 5090 report
2× RTX 30908-bit160K168 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on 2× RTX 3090 report
2× RTX 40908-bit160K178 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit160K216 tok/s0.2 sPDF: DeepSeek Coder V2 Lite Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit160K172 tok/s0.2 sPDF: DeepSeek Coder V2 Lite Instruct on RTX 6000 Ada report
L40S8-bit160K159 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on L40S report
2× RTX 50908-bit160K258 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 2× RTX 5090 report
4× RTX 30908-bit160K168 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit160K258 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB4-bit24K50 tok/s11 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit88K39 tok/s11 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit136K31 tok/s11 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit88K74 tok/s5.4 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit160K61 tok/s5.4 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit160K61 tok/s5.4 sPDF: DeepSeek Coder V2 Lite Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit160K94 tok/s2.7 sPDF: DeepSeek Coder V2 Lite Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit160K101 tok/s0.8 sPDF: DeepSeek Coder V2 Lite Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit160K110 tok/s1.5 sPDF: DeepSeek Coder V2 Lite Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit160K110 tok/s1.5 sPDF: DeepSeek Coder V2 Lite Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit160K110 tok/s1.5 sPDF: DeepSeek Coder V2 Lite Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit160K135 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit160K48 tok/s1.5 sPDF: DeepSeek Coder V2 Lite Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit160K48 tok/s1.5 sPDF: DeepSeek Coder V2 Lite Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit160K66 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on RTX Spark report
DGX Spark8-bit160K61 tok/s0.5 sPDF: DeepSeek Coder V2 Lite Instruct on DGX Spark report
Dual DGX Spark8-bit160K61 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on Dual DGX Spark report
DGX Station GB3008-bit160K457 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit160K258 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit160K258 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit160K258 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit160K278 tok/s0.3 sPDF: DeepSeek Coder V2 Lite Instruct on A100 80 GB SXM report
H100 SXM8-bit160K354 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on H100 SXM report
4× H200 NVL8-bit160K406 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct on 4× H200 NVL report
8× H200 NVL8-bit160K406 tok/s0.1 sPDF: DeepSeek Coder V2 Lite Instruct 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 lmstudio-community/DeepSeek-Coder-V2-Lite-Instruct-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-Coder-V2-Lite-Instruct-4bit

Questions

How much VRAM does DeepSeek Coder V2 Lite Instruct need?

DeepSeek Coder V2 Lite Instruct needs 8.9 GB for its weights at 4-bit, 9.8 GB with an 8K context and 10.5 GB with 32K. At 8-bit, 16.5 GB.

What is the minimum VRAM for DeepSeek Coder V2 Lite Instruct?

About 8.1 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 3080 10 GB.

How big is DeepSeek Coder V2 Lite Instruct?

15.7B parameters (2.7B active per token): 29.3 GB of weights at 16-bit, 8.9 GB at 4-bit.

Can DeepSeek Coder V2 Lite Instruct run on an RTX 4090?

Yes. At 8-bit it holds 160K of context and writes about 178 tokens a second.

More models