mistralai · audio asr

Voxtral-Mini-4B-Realtime-2602

mistralai/Voxtral-Mini-4B-Realtime-2602

Voxtral-Mini-4B-Realtime-2602 at Q4_K_M is exactly 2,830,493,984 bytes (2.64 GiB / 2.83 GB) — an effective 5.112 bits per weight, not the nominal 4.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
4.4B
Architecture
voxtral_realtime
26 layers
Context
131,072
native (config.json)
License
apache-2.0

Shipped quantizations

exact bytes, summed from published files
QuantSizeExact bytesEffective bpwTensorsPublisher
Q4_K2.35 GiB2,524,137,2164.559cstr
Q4_K_M2.64 GiB2,830,493,9845.112714handy-computer
Q5_K_M3.06 GiB3,281,439,0085.926714handy-computer
Q6_K3.41 GiB3,661,018,9126.612714handy-computer
Q8_04.41 GiB4,731,791,6488.546714handy-computer
Q8_04.41 GiB4,733,486,8488.549714cstr
BF168.26 GiB8,868,301,08816.016handy-computer
F168.27 GiB8,879,114,52816.036714handy-computer

KV cache by context

unresolved

This model declares a 8,192-token sliding window, but we could not establish which layers use it. Its architecture publishes the layout as a per-layer array inside the model file rather than as a period in config.json, and we have not yet ingested that array.

A flat context × layers × heads figure would be substantially too high, so we are not showing one. This is tracked as a known gap rather than filled with a guess.

Measured

published by a third party, attributed below
MetricValueWhat it means
rtf105.1
RTFx105.1Higher is better — audio seconds processed per second of compute.
Word error rate13.34%Lower is better — the share of words transcribed incorrectly.
Word error rate2.60%Lower is better — the share of words transcribed incorrectly.
Word error rate15.91%Lower is better — the share of words transcribed incorrectly.
Word error rate8.86%Lower is better — the share of words transcribed incorrectly.
Word error rate2.23%Lower is better — the share of words transcribed incorrectly.
Word error rate8.00%Lower is better — the share of words transcribed incorrectly.
Word error rate4.92%Lower is better — the share of words transcribed incorrectly.
Word error rate8.80%Lower is better — the share of words transcribed incorrectly.
Word error rate6.42%Lower is better — the share of words transcribed incorrectly.
Word error rate1.61%Lower is better — the share of words transcribed incorrectly.
Word error rate11.46%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

Will it run on your card?

full quant x context sweep

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 2.32 GiB. The real file is 2.64 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
26
Attention heads
32
KV heads
8
Head dim
128
Hidden size
3072
Vocab
131,072
Sliding window
8192
SWA period
MLA
no
Experts
Experts per token
use_sliding_window

Questions people ask

How much VRAM does Voxtral-Mini-4B-Realtime-2602 need?
Q4_K_M is exactly 2,830,493,984 bytes (2.64 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Which quantization of Voxtral-Mini-4B-Realtime-2602 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.