How much VRAM does Voxtral-Mini-3B-2507 need?

Q4_K_M is exactly 2,473,002,304 bytes (2.30 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Voxtral-Mini-3B-2507's KV cache?

3.75 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-Mini-3B-2507

mistralai/Voxtral-Mini-3B-2507

Voxtral-Mini-3B-2507 at Q4_K_M is exactly 2,473,002,304 bytes (2.30 GiB / 2.47 GB) — an effective 4.231 bits per weight, not the nominal 4. Its KV cache at 32K is 3.75 GiB.

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

Parameters

4.7B

Architecture

voxtral

30 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_M	1.45 GiB	1,559,168,320	2.667	—	bartowski
Q2_K	1.55 GiB	1,661,551,936	2.842	—	bartowski
IQ3_XXS	1.58 GiB	1,692,468,544	2.895	—	bartowski
IQ3_XS	1.70 GiB	1,827,341,632	3.126	—	bartowski
Q3_K_S	1.76 GiB	1,891,190,080	3.235	—	bartowski
IQ3_M	1.83 GiB	1,963,885,888	3.360	—	bartowski
Q2_K_L	1.91 GiB	2,054,767,936	3.515	—	bartowski
Q3_K_M	1.92 GiB	2,058,061,120	3.521	—	bartowski
Q3_K_L	2.06 GiB	2,207,483,200	3.776	—	bartowski
IQ4_XS	2.12 GiB	2,272,707,904	3.888	—	bartowski
Q4_0	2.21 GiB	2,375,533,888	4.064	—	bartowski
Q4_K_S	2.22 GiB	2,381,825,344	4.075	—	bartowski
IQ4_NL	2.22 GiB	2,382,611,776	4.076	—	bartowski
Q4_K_M	2.30 GiB	2,473,002,304	4.231	—	bartowski
Q4_1	2.42 GiB	2,596,521,280	4.442	—	bartowski
Q4_K	2.47 GiB	2,652,639,968	4.538	—	cstr
Q4_K_L	2.58 GiB	2,771,846,464	4.742	—	bartowski
Q5_K_S	2.63 GiB	2,822,227,264	4.828	—	bartowski
Q5_K_M	2.68 GiB	2,874,869,056	4.918	—	bartowski
Q4_K_M	2.78 GiB	2,984,721,056	5.106	—	handy-computer
Q5_K_L	2.91 GiB	3,123,381,568	5.343	—	bartowski
Q6_K	3.08 GiB	3,301,852,480	5.649	—	bartowski
Q5_K_M	3.23 GiB	3,464,182,432	5.926	—	handy-computer
Q6_K_L	3.26 GiB	3,496,887,616	5.982	—	bartowski
Q6_K	3.60 GiB	3,869,489,824	6.620	—	handy-computer
Q8_0	3.98 GiB	4,273,980,736	7.312	—	bartowski
Q8_0	4.64 GiB	4,986,770,144	8.531	—	cstr
Q8_0	4.66 GiB	5,000,084,128	8.554	—	handy-computer
BF16	7.49 GiB	8,037,057,568	13.749	—	bartowski
BF16	8.72 GiB	9,365,764,768	16.023	—	handy-computer
F16	8.73 GiB	9,376,578,208	16.041	—	handy-computer

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.47 GiB	0.47 GiB	—	30 / 0 / 0
8,192	0.94 GiB	0.94 GiB	—	30 / 0 / 0
16,384	1.88 GiB	1.88 GiB	—	30 / 0 / 0
32,768	3.75 GiB	3.75 GiB	—	30 / 0 / 0
65,536	7.50 GiB	7.50 GiB	—	30 / 0 / 0
131,072	15.00 GiB	15.00 GiB	—	30 / 0 / 0

Measured

published by a third party, attributed below

Metric	Value	What it means
rtf	179.7
RTFx	179.7	Higher is better — audio seconds processed per second of compute.
Word error rate	13.56%	Lower is better — the share of words transcribed incorrectly.
Word error rate	10.36%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.70%	Lower is better — the share of words transcribed incorrectly.
Word error rate	1.47%	Lower is better — the share of words transcribed incorrectly.
Word error rate	3.62%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.15%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.22%	Lower is better — the share of words transcribed incorrectly.
Word error rate	15.04%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.75%	Lower is better — the share of words transcribed incorrectly.
Word error rate	6.77%	Lower is better — the share of words transcribed incorrectly.
Word error rate	6.01%	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 cohere-transcribe-03-2026 vs parakeet-tdt-0.6b-v3 vs whisper-medium vs Voxtral-Mini-4B-Realtime-2602 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 2.45 GiB. The real file is 2.30 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

3072

Vocab

131,072

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Voxtral-Mini-3B-2507 need?: Q4_K_M is exactly 2,473,002,304 bytes (2.30 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-Mini-3B-2507's KV cache?: 3.75 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-Mini-3B-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.