How much VRAM does Voxtral-Small-24B-2507 need?

Q4_K_M is exactly 14,302,261,728 bytes (13.32 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Voxtral-Small-24B-2507's KV cache?

5.00 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.

mistralai · audio asr

Voxtral-Small-24B-2507

mistralai/Voxtral-Small-24B-2507

Voxtral-Small-24B-2507 at Q4_K_M is exactly 14,302,261,728 bytes (13.32 GiB / 14.30 GB) — an effective 4.716 bits per weight, not the nominal 4. Its KV cache at 32K is 5.00 GiB.

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

Parameters

24.3B

Architecture

voxtral

40 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XS	6.71 GiB	7,207,566,272	2.377	—	bartowski
IQ2_S	6.97 GiB	7,478,885,312	2.466	—	bartowski
IQ2_M	7.56 GiB	8,114,584,512	2.676	—	bartowski
Q2_K	8.28 GiB	8,890,858,432	2.932	—	bartowski
IQ3_XXS	8.64 GiB	9,281,125,312	3.060	—	bartowski
Q2_K_L	8.89 GiB	9,546,218,432	3.148	—	bartowski
IQ3_XS	9.23 GiB	9,907,649,472	3.267	—	bartowski
Q3_K_S	9.69 GiB	10,400,807,872	3.429	—	bartowski
IQ3_M	9.92 GiB	10,651,483,072	3.512	—	bartowski
Q3_K_M	10.69 GiB	11,474,615,232	3.784	363	bartowski
Q3_K_L	11.55 GiB	12,401,294,272	4.089	—	bartowski
IQ4_XS	11.88 GiB	12,759,448,512	4.207	363	bartowski
IQ4_NL	12.54 GiB	13,468,548,032	4.441	—	bartowski
Q4_0	12.57 GiB	13,494,762,432	4.450	363	bartowski
Q4_K_S	12.62 GiB	13,549,812,672	4.468	—	bartowski
Q4_K_M	13.32 GiB	14,302,261,728	4.716	—	handy-computer
Q4_K_M	13.35 GiB	14,334,442,432	4.727	363	bartowski
Q4_K_L	13.81 GiB	14,832,516,032	4.891	—	bartowski
Q4_1	13.85 GiB	14,873,639,872	4.904	—	bartowski
Q5_K_S	15.19 GiB	16,304,946,112	5.376	—	bartowski
Q5_K_M	15.61 GiB	16,764,517,312	5.528	363	bartowski
Q5_K_M	15.96 GiB	17,138,659,808	5.651	—	handy-computer
Q5_K_L	16.00 GiB	17,178,704,832	5.664	—	bartowski
Q6_K	18.02 GiB	19,346,471,872	6.379	363	bartowski
Q6_K_L	18.32 GiB	19,671,530,432	6.486	—	bartowski
Q6_K	18.57 GiB	19,936,473,568	6.574	—	handy-computer
Q8_0	23.33 GiB	25,055,312,832	8.262	363	bartowski
Q8_0	24.04 GiB	25,810,383,328	8.511	—	handy-computer
BF16	43.92 GiB	47,154,051,744	15.548	—	bartowski
BF16	45.20 GiB	48,537,285,088	16.005	—	handy-computer
F16	45.21 GiB	48,548,098,528	16.008	854	handy-computer

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.63 GiB	0.63 GiB	—	40 / 0 / 0
8,192	1.25 GiB	1.25 GiB	—	40 / 0 / 0
16,384	2.50 GiB	2.50 GiB	—	40 / 0 / 0
32,768	5.00 GiB	5.00 GiB	—	40 / 0 / 0
65,536	10.00 GiB	10.00 GiB	—	40 / 0 / 0
131,072	20.00 GiB	20.00 GiB	—	40 / 0 / 0

Measured

published by a third party, attributed below

Metric	Value	What it means
rtf	100.1
RTFx	100.1	Higher is better — audio seconds processed per second of compute.
Word error rate	1.84%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.00%	Lower is better — the share of words transcribed incorrectly.
Word error rate	14.32%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.37%	Lower is better — the share of words transcribed incorrectly.
Word error rate	10.16%	Lower is better — the share of words transcribed incorrectly.
Word error rate	13.18%	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	1.23%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.33%	Lower is better — the share of words transcribed incorrectly.
Word error rate	6.60%	Lower is better — the share of words transcribed incorrectly.
Word error rate	5.65%	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 cohere-transcribe-03-2026 vs whisper-medium vs Voxtral-Mini-4B-Realtime-2602 vs whisper-large-v3 vs Qwen3-ASR-1.7B vs canary-1b-v2 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 12.71 GiB. The real file is 13.32 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

5120

Vocab

131,072

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Voxtral-Small-24B-2507 need?: Q4_K_M is exactly 14,302,261,728 bytes (13.32 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Voxtral-Small-24B-2507's KV cache?: 5.00 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 Voxtral-Small-24B-2507 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.