How much VRAM does deepseek-llm-67b-base need?

Q4_K_M is exactly 40,439,686,592 bytes (37.66 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is deepseek-llm-67b-base's KV cache?

11.88 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

deepseek-ai · text

deepseek-llm-67b-base

deepseek-ai/deepseek-llm-67b-base

deepseek-llm-67b-base at Q4_K_M is exactly 40,439,686,592 bytes (37.66 GiB / 40.44 GB) — an effective 4.798 bits per weight, not the nominal 4. Its KV cache at 32K is 11.88 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

67.4B

Architecture

llama

95 layers

Context

4,096

native (config.json)

License

other

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	13.65 GiB	14,658,184,672	1.739	—	mradermacher
I1-IQ1_M	14.89 GiB	15,984,502,240	1.897	—	mradermacher
I1-IQ2_XXS	16.95 GiB	18,195,031,520	2.159	—	mradermacher
I1-IQ2_XS	18.78 GiB	20,162,684,384	2.392	—	mradermacher
I1-IQ2_S	19.88 GiB	21,345,904,096	2.533	—	mradermacher
I1-IQ2_M	21.53 GiB	23,114,327,520	2.743	—	mradermacher
I1-Q2_K	23.40 GiB	25,121,301,984	2.981	—	mradermacher
I1-IQ3_XXS	24.32 GiB	26,109,912,544	3.098	—	mradermacher
I1-IQ3_XS	25.95 GiB	27,861,624,288	3.306	—	mradermacher
Q2_K	26.54 GiB	28,495,750,592	3.381	—	TheBloke
Q3_K_S	27.30 GiB	29,309,380,032	3.478	—	TheBloke
I1-Q3_K_S	27.30 GiB	29,309,380,064	3.478	—	mradermacher
I1-IQ3_S	27.40 GiB	29,415,220,704	3.490	—	mradermacher
I1-IQ3_M	28.43 GiB	30,525,433,312	3.622	—	mradermacher
Q3_K_M	30.35 GiB	32,584,705,472	3.866	—	TheBloke
I1-Q3_K_M	30.41 GiB	32,652,338,656	3.874	—	mradermacher
Q3_K_L	33.13 GiB	35,575,768,512	4.221	—	TheBloke
I1-Q3_K_L	33.13 GiB	35,575,768,544	4.221	—	mradermacher
I1-IQ4_XS	33.71 GiB	36,196,394,464	4.295	—	mradermacher
Q4_0	35.53 GiB	38,152,644,032	4.527	—	TheBloke
Q4_K_S	35.62 GiB	38,247,015,872	4.538	—	TheBloke
I1-Q4_0	35.65 GiB	38,276,638,176	4.542	—	mradermacher
I1-Q4_K_S	35.77 GiB	38,404,826,592	4.557	—	mradermacher
Q4_K_M	37.66 GiB	40,439,686,592	4.798	—	TheBloke
I1-Q4_K_M	37.66 GiB	40,439,686,624	4.798	—	mradermacher
Q5_K_S	43.28 GiB	46,475,716,032	5.514	—	TheBloke
Q5_0	43.28 GiB	46,475,716,032	5.514	—	TheBloke
I1-Q5_K_S	43.28 GiB	46,475,716,064	5.514	—	mradermacher
Q5_K_M	44.38 GiB	47,653,889,472	5.654	—	TheBloke
I1-Q5_K_M	44.38 GiB	47,653,889,504	5.654	—	mradermacher

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.48 GiB	1.48 GiB	—	95 / 0 / 0
8,192	2.97 GiB	2.97 GiB	—	95 / 0 / 0
16,384	5.94 GiB	5.94 GiB	—	95 / 0 / 0
32,768	11.88 GiB	11.88 GiB	—	95 / 0 / 0
65,536	23.75 GiB	23.75 GiB	—	95 / 0 / 0
131,072	47.50 GiB	47.50 GiB	—	95 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 35.32 GiB. The real file is 37.66 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

8192

Vocab

102,400

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does deepseek-llm-67b-base need?: Q4_K_M is exactly 40,439,686,592 bytes (37.66 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is deepseek-llm-67b-base's KV cache?: 11.88 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of deepseek-llm-67b-base should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.