How much VRAM does whisper-large-v3-turbo need?

Q4_K_M is exactly 536,069,728 bytes (0.50 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

openai · audio asr

whisper-large-v3-turbo

openai/whisper-large-v3-turbo

whisper-large-v3-turbo at Q4_K_M is exactly 536,069,728 bytes (0.50 GiB / 0.54 GB) — an effective 5.302 bits per weight, not the nominal 4.

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

Parameters

809M

Architecture

whisper

32 layers

Context

—

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	0.27 GiB	286,400,435	2.833	—	oxide-lab
Q2_K	0.27 GiB	286,400,435	2.833	—	Xviers
Q3_K	0.34 GiB	367,690,215	3.636	—	oxide-lab
Q3_K	0.34 GiB	367,690,215	3.636	—	Xviers
Q4_0	0.43 GiB	462,920,256	4.578	588	Xviers
Q4_0	0.43 GiB	462,920,256	4.578	—	oxide-lab
Q4_0	0.44 GiB	473,992,235	4.688	—	oxide-lab
Q4_K	0.44 GiB	473,992,235	4.688	—	Xviers
Q4_0	0.44 GiB	473,992,235	4.688	—	Xviers
Q4_K	0.44 GiB	473,992,235	4.688	—	oxide-lab
Q4_1	0.48 GiB	513,132,192	5.075	—	Xviers
Q4_1	0.48 GiB	513,132,192	5.075	—	oxide-lab
Q4_1	0.49 GiB	524,016,715	5.183	—	Xviers
Q4_1	0.49 GiB	524,016,715	5.183	—	oxide-lab
Q4_K_M	0.50 GiB	536,069,728	5.302	589	handy-computer
Q5_K	0.53 GiB	574,041,195	5.677	—	Xviers
Q5_0	0.53 GiB	574,041,195	5.677	—	Xviers
Q5_0	0.53 GiB	574,041,195	5.677	—	oxide-lab
Q5_K	0.53 GiB	574,041,195	5.677	—	oxide-lab
Q5_K_M	0.58 GiB	619,628,128	6.128	589	handy-computer
Q5_1	0.58 GiB	624,065,675	6.172	—	oxide-lab
Q5_1	0.58 GiB	624,065,675	6.172	—	Xviers
Q6_K	0.63 GiB	680,343,215	6.729	—	Xviers
Q6_K	0.63 GiB	680,343,215	6.729	—	oxide-lab
Q6_K	0.64 GiB	692,536,928	6.849	589	handy-computer
Q8_0	0.81 GiB	864,655,936	8.552	588	oxide-lab
Q8_0	0.81 GiB	864,655,936	8.552	—	Xviers
Q8_0	0.81 GiB	874,188,075	8.646	—	Xviers
Q8_0	0.81 GiB	874,188,075	8.646	—	oxide-lab
Q8_0	0.83 GiB	886,381,760	8.767	—	handy-computer
F16	1.51 GiB	1,625,935,520	16.081	589	handy-computer

Measured

published by a third party, attributed below

Metric	Value	What it means
rtf	782.6
RTFx	782.6	Higher is better — audio seconds processed per second of compute.
Word error rate	11.22%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.52%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.13%	Lower is better — the share of words transcribed incorrectly.
Word error rate	15.16%	Lower is better — the share of words transcribed incorrectly.
Word error rate	7.02%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.79%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.47%	Lower is better — the share of words transcribed incorrectly.
Word error rate	13.87%	Lower is better — the share of words transcribed incorrectly.
Word error rate	3.70%	Lower is better — the share of words transcribed incorrectly.
Word error rate	11.07%	Lower is better — the share of words transcribed incorrectly.
Word error rate	7.01%	Lower is better — the share of words transcribed incorrectly.

Benchmarked· by open-asr-leaderboard-english-short-latest

RTFx measured by the Open ASR Leaderboard on a single datacenter GPU at a large batch size. It ranks models against each other; it says nothing about throughput on consumer hardware. We reproduce these figures with attribution; they are not ours and we have not verified the runs. Source: open-asr-leaderboard-english-short-latest.

Compare with

same modality, comparable size

vs nemotron-3.5-asr-streaming-0.6b vs parakeet-unified-en-0.6b vs parakeet-tdt-0.6b-v3 vs whisper-medium vs whisper-large-v3 vs canary-180m-flash 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 0.42 GiB. The real file is 0.50 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

—

Vocab

51,866

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does whisper-large-v3-turbo need?: Q4_K_M is exactly 536,069,728 bytes (0.50 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Which quantization of whisper-large-v3-turbo 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.