How much VRAM does DeepSeek-OCR-2 need?

Q4_K_M is exactly 1,950,326,688 bytes (1.82 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is DeepSeek-OCR-2's KV cache?

0.94 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.

Is DeepSeek-OCR-2 a mixture-of-experts model?

Yes — 64 experts with 6 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

deepseek-ai · text · mixture of experts

DeepSeek-OCR-2

deepseek-ai/DeepSeek-OCR-2

DeepSeek-OCR-2 at Q4_K_M is exactly 1,950,326,688 bytes (1.82 GiB / 1.95 GB) — an effective 4.604 bits per weight, not the nominal 4. Its KV cache at 32K is 0.94 GiB.

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

Parameters

3.4B

total, not active

Architecture

deepseek2-ocr

12 layers

Context

8,192

native (config.json)

License

—

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ3_M	1.35 GiB	1,448,949,408	3.420	—	SandLogicTechnologies
IQ4_XS	1.53 GiB	1,642,540,448	3.877	—	SandLogicTechnologies
IQ4_NL	1.59 GiB	1,703,545,248	4.021	—	SandLogicTechnologies
Q4_K_M	1.82 GiB	1,950,326,688	4.604	—	sabafallah
Q8_0	2.91 GiB	3,126,139,808	7.379	—	JamePeng2023
Q8_0	2.91 GiB	3,126,139,808	7.379	—	sabafallah
BF16	5.47 GiB	5,876,578,208	13.872	—	sabafallah
BF16	5.47 GiB	5,876,578,208	13.872	—	JamePeng2023

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.12 GiB	0.12 GiB	—	12 / 0 / 0
8,192	0.23 GiB	0.23 GiB	—	12 / 0 / 0
16,384	0.47 GiB	0.47 GiB	—	12 / 0 / 0
32,768	0.94 GiB	0.94 GiB	—	12 / 0 / 0
65,536	1.88 GiB	1.88 GiB	—	12 / 0 / 0
131,072	3.75 GiB	3.75 GiB	—	12 / 0 / 0

Compare with

same modality, comparable size

vs gemma-4-E4B-it vs Qwen3-4B vs Qwen3-8B vs HyperCLOVAX-SEED-Text-Instruct-1.5B vs Llama-3.2-1B-Instruct vs Qwen3-VL-8B-Instruct-abliterated-v1 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 1.78 GiB. The real file is 1.82 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

1280

Vocab

129,280

Sliding window

none

SWA period

—

MLA

Experts

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does DeepSeek-OCR-2 need?: Q4_K_M is exactly 1,950,326,688 bytes (1.82 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-OCR-2's KV cache?: 0.94 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.
Is DeepSeek-OCR-2 a mixture-of-experts model?: Yes — 64 experts, 6 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of DeepSeek-OCR-2 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.