DeepSeek

Deepseek Llm 67B

DeepSeek·67B·DeepSeek·4K ctx·2026-08

Check on my hardwareAdd to compare →Ollama →Hugging Face →

Model size→Hardware requirement→Actual fit

Deepseek Llm 67B needs ~39.2–46.1 GB of VRAM depending on quantization. The bar is what it needs; the marker is the card's VRAM. Speed figures below are planning estimates.

8 GB

RTX 4060 · 3060 Ti

39.2
8
Q4_K_M · Est. ~1 tok/sWon't run

12 GB

RTX 4070 · 3060 12G

39.2
12
Q4_K_M · Est. ~1 tok/sWon't run

16 GB

RTX 4080 · M-series 16G

39.2
16
Q4_K_M · Est. ~1 tok/sWon't run

24 GB

RTX 4090 · 3090

39.2
24
Q4_K_M · Est. ~7 tok/sSlow · CPU offload

48 GB

A6000 · 2×24G

46.1
48
Q5_K_M · Est. ~31 tok/sComfortable

02  — Quantizations

QuantFile sizeEst. VRAM
Q4_K_M38 GB39.2 GB
Q5_K_M44.9 GB46.1 GB

03  — Runtimes & use

llama.cpp · Ollama · LM Studio

General · Max intelligence

Memory and speed figures are planning estimates, not measured benchmarks. Memory includes a fixed allowance for KV cache and runtime overhead. Long-context prompts push VRAM use above the numbers shown here.