handy models

speech-to-text benchmarks for transcribe.cpp, the engine behind Handy

model detail

Where this model stands against the best model in each language at Q8_0, its speed on an AMD Ryzen 4750U laptop, and what the card says. Add a second model to compare them head to head.

model Voxtral-Mini-4B-Realtime-2602
AMD 4750U GPU0.8×
AMD 4750U CPU0.6×
Q8_0 size4.41 GB
frontier · AMD 4750U GPU2 of 13

verdict: Best in 2 of 13 languages, and tied with the leader in 4 more (overlapping 95% CIs). Weakest relative to the field in Japanese (CER) (2.3× the leader's error), Hindi (2.0×), and English (1.9×). Slower than realtime on the AMD 4750U GPU (0.8×) and AMD 4750U CPU (0.6×).

license: Apache-2.0 · card languages: 13 · streaminglang detect

standing by language · 13 benchmarked

Each bar compares this model with the best other model in that language, on a log scale: left of center it is ahead, right of center behind, and the other model is named on the row. The two error rates use the legend colors: this model first, the other model second. Faded bars are ties, where the 95% confidence intervals overlap. Sorted strongest standing first. Hover a row for full details.

Voxtral-Mini-4B-Realtime-2602 error best other model error tie: the 95% CIs overlap

quantization · 13 usable languages at Q5_K_M

Q5_K_M is within noise on every one of 13 usable languages: 69% of the Q8_0 file for no measurable accuracy cost.

Q5_K_M file69% of Q8_03.06 GB against 4.41 GB
median error change+0.2%over 13 usable languages
measurably worse0 of 13above the Q8_0 95% confidence interval
worst usable languageJapanese (CER)5.44 → 5.78 (+0.34 pts)
measurably worse than Q8_0: above its 95% confidence interval within noise typical noise of one measurement
per-language numbers (13)
Japanese (CER) 5.444.94–5.95 5.78+0.34 +6.2%
English 6.275.38–7.31 6.57+0.30 +4.8%
Korean (CER) 5.274.30–6.31 5.36+0.09 +1.7%
Arabic 14.5313.17–16.13 14.77+0.24 +1.7%
Russian 5.414.95–5.88 5.47+0.06 +1.1%
German 4.243.78–4.73 4.27+0.03 +0.7%
French 6.295.67–6.88 6.30+0.01 +0.2%
Portuguese 3.873.43–4.38 3.87+0.00 +0.0%
Dutch 6.095.45–6.73 6.07−0.02 −0.3%
Spanish 2.672.38–2.96 2.66−0.01 −0.4%
Italian 2.251.99–2.53 2.24−0.01 −0.4%
Hindi 14.4413.36–15.68 14.36−0.08 −0.6%
Mandarin Chinese (CER) 10.419.32–11.52 8.57−1.84 −17.7%

speed · all 73 models on the AMD 4750U GPU

× realtime, log scale. Left of the realtime marker, transcribing takes longer than the recording. Hover a dot for the model.

compare with

Voxtral-Mini-3B-2507same familyVoxtral-Small-24B-2507same familywhisper-large-v3ahead of it in 8 of 13 shared languagescohere-transcribe-03-2026closest coverage, 11 languages in common