How much VRAM does DeepSeek LLM 67B need?

DeepSeek LLM 67B (67B parameters) requires approximately 61.7 GB of memory with Q4_K_M quantization.

What is the best quantization for DeepSeek LLM 67B?

The recommended quantization for DeepSeek LLM 67B is Q4_K_M, which balances quality and memory efficiency.

Can it run?

Yes, NVIDIA A100 80GB can run DeepSeek LLM 67B with a C grade (Runs well). Expected decode speed: 35.9 tok/s.

CUsable

Runs well

Using Q4_K_M in vLLM

Capabilities:

Fit status

Runs well

Decode

35.9 tok/s

TTFT

5390 ms

Safe context

Memory

61.7 GB / 80.0 GB

Weights40.9 GB

KV Cache10.5 GB

Runtime2.4 GB

Headroom8.0 GB

Workload	Grade	Fit	Decode	TTFT	Context
Agentic Coding	C	Tight fit	41.9 tok/s	6720 ms	4K
Chat	C	Runs well	41.9 tok/s	2520 ms	4K
Coding	C	Runs well	35.9 tok/s	5390 ms	4K
RAG	C	Tight fit	41.9 tok/s	8399 ms	4K
Reasoning	C	Runs well	41.9 tok/s	5460 ms	4K

How DeepSeek LLM 67B (67B params) fits at each quantization level on NVIDIA A100 80GB (80.0 GB usable).

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	26.1 GB	Low	D36
Q3_K_S	3	32.8 GB	Low	D38
NVFP4	4

Ollama

ollama run deepseek-llm-67b

HuggingFace

huggingface-cli download deepseek-llm-67b

Upgrade options

NVIDIA GH200 96GBBudget pick

B79.3 tok/s decode

NVIDIA H20 96GBBiggest leap

B79.3 tok/s decode