How much VRAM does bge-m3 need?

Q8_0 is exactly 634,553,760 bytes (0.59 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

BAAI · embedding

bge-m3

BAAI/bge-m3

bge-m3 at Q8_0 is exactly 634,553,760 bytes (0.59 GiB / 0.63 GB) — an effective 8.958 bits per weight, not the nominal 8.

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

Parameters

567M

Architecture

bert

24 layers

Context

8,194

native (config.json)

License

mit

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q8_0	0.59 GiB	634,553,760	8.958	—	ggml-org

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 jina-embeddings-v5-text-small vs embeddinggemma-300m-qat-q8_0-unquantized vs KaLM-embedding-multilingual-mini-instruct-v2.5 vs nomic-embed-text-v1.5 vs jina-embeddings-v5-text-nano vs nomic-embed-text-v2-moe All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q8_0 at roughly 0.30 GiB. The real file is 0.59 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

1024

Vocab

250,002

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does bge-m3 need?: Q8_0 is exactly 634,553,760 bytes (0.59 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Which quantization of bge-m3 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.