Skip to content
localmodel.run

Text model · DeepSeek-V2

DeepSeek-V2-Lite requirements

DeepSeek-V2 family · 16B params (Mixture-of-Experts, 2.4B active) · released May 2024 · 537.3K Ollama pulls. Minimum to run at Q4_K_M: Nvidia GeForce RTX 4060 Ti (16GB).

LicenseDeepSeek License· Commercial OK↓ 247.1K/mo♥ 179on HuggingFace
Q4_K_M
10.4 GB
Q8_0
16.8 GB
Total @ Q4 (4k)
~12.2 GB
Context
32 k

Quantization sizes

GGUF quantson disk
QuantizationSize on disk
Q2_K6.7 GB est
Q3_K_M7.8 GB est
Q4_K_M (default)10.4 GB
Q5_K_M11.4 GB est
Q6_K13.1 GB est
Q8_016.8 GB
FP1632 GB est

Lower quant = smaller and faster, slightly lower quality. Q4_K_M is the common default.

Run it

Ollama
$ ollama run deepseek-v2:16b
llama.cpp
$ llama-cli -hf mradermacher/DeepSeek-V2-Lite-GGUF:Q4_K_M
LM Studio
$ lms get mradermacher/DeepSeek-V2-Lite-GGUF

Which devices can run DeepSeek-V2-Lite?

FAQ

How much VRAM or RAM does DeepSeek-V2-Lite need?

At Q4_K_M, DeepSeek-V2-Lite needs about 12.2 GB (weights ~10.4 GB + KV cache + overhead) at a 4k context. At Q8_0 budget ~18.6 GB.

Can DeepSeek-V2-Lite run on a laptop?

DeepSeek-V2-Lite is large; you need a 24 GB+ GPU or a 32-48 GB Mac at Q4_K_M.

Is DeepSeek-V2-Lite cheaper to run because it is a MoE model?

It is faster, not lighter. DeepSeek-V2-Lite activates only 2.4B of 16B params per token (so it runs quickly), but all experts must stay in memory, so it still needs memory for the full 16B.

Can I use DeepSeek-V2-Lite commercially?

Yes. DeepSeek-V2-Lite is licensed DeepSeek License, which permits commercial use.

MoE: 16B total / 2.4B active (2 shared + 64 routed experts per layer, 6 activated per token). Released 2024-05-16. Q4_K_M=10.4GB, Q8_0=16.8GB from mradermacher HF GGUF repo (bartowski does not host this model). Ollama deepseek-v2:16b=8.9GB uses different quantization. Context 32K native. Despite 16B total params, inference speed is close to 2.4B dense models.

Sources

Memory figures are estimates. See methodology.