How much VRAM does all-MiniLM-L6-v2 need?

Q4_K_M is exactly 20,999,104 bytes (0.02 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

sentence-transformers · embedding

all-MiniLM-L6-v2

sentence-transformers/all-MiniLM-L6-v2

all-MiniLM-L6-v2 at Q4_K_M is exactly 20,999,104 bytes (0.02 GiB / 0.02 GB) — an effective 7.396 bits per weight, not the nominal 4.

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

Parameters

23M

Architecture

bert

6 layers

Context

512

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	0.02 GiB	19,229,632	6.773	—	leliuga
Q3_K_S	0.02 GiB	19,229,632	6.773	—	leliuga
Q3_K_S	0.02 GiB	19,229,632	6.773	—	second-state
Q2_K	0.02 GiB	19,229,632	6.773	—	second-state
Q4_0	0.02 GiB	19,699,648	6.938	—	leliuga
Q4_0	0.02 GiB	19,699,648	6.938	101	second-state
Q3_K_M	0.02 GiB	19,939,264	7.023	101	leliuga
Q3_K_M	0.02 GiB	19,939,264	7.023	—	second-state
Q3_K_L	0.02 GiB	20,473,792	7.211	—	leliuga
Q3_K_L	0.02 GiB	20,473,792	7.211	—	second-state
Q4_K_S	0.02 GiB	20,694,976	7.289	—	leliuga
Q4_K_S	0.02 GiB	20,694,976	7.289	—	second-state
Q4_K_M	0.02 GiB	20,999,104	7.396	—	leliuga
Q4_K_M	0.02 GiB	20,999,104	7.396	101	second-state
Q5_0	0.02 GiB	21,026,752	7.406	—	second-state
Q5_0	0.02 GiB	21,026,752	7.406	—	leliuga
Q5_K_S	0.02 GiB	21,469,120	7.562	—	leliuga
Q5_K_S	0.02 GiB	21,469,120	7.562	—	second-state
Q5_K_M	0.02 GiB	21,717,952	7.649	—	second-state
Q5_K_M	0.02 GiB	21,717,952	7.649	101	leliuga
Q6_K	0.02 GiB	24,150,976	8.506	101	second-state
Q6_K	0.02 GiB	24,150,976	8.506	—	leliuga
Q8_0	0.02 GiB	25,008,064	8.808	101	second-state
Q8_0	0.02 GiB	25,008,064	8.808	—	leliuga
Q8_0	0.02 GiB	25,282,176	8.905	—	sudomoniker
F16	0.04 GiB	45,949,216	16.184	—	second-state
F16	0.04 GiB	45,949,216	16.184	101	leliuga
F32	0.08 GiB	91,016,992	32.057	—	leliuga

No KV cache

architectural, not a gap in our data

This architecture allocates no KV cache. Encoder and embedding models process their input in one pass rather than generating token by token, so there is nothing to carry forward between steps and memory does not grow with context. Its footprint is the weights plus a working buffer, and that is the whole story.

Compare with

same modality, comparable size

vs mxbai-embed-xsmall-v1 vs jina-reranker-v1-tiny-en vs gte-small 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 0.01 GiB. The real file is 0.02 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

Hidden size

384

Vocab

30,522

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does all-MiniLM-L6-v2 need?: Q4_K_M is exactly 20,999,104 bytes (0.02 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Which quantization of all-MiniLM-L6-v2 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.