Qwen3-ASR-0.6B

+ contribute

Qwen3-ASR-0.6B is an open-weights speech-to-text model from Alibaba (Qwen), released January 2026. Its published size is 0.94B parameters. Licence: apache-2.0.

Catalogue identity notes

Open-weights Qwen/Qwen3-ASR-0.6B (Apache-2.0, released 2026-01-28), the smaller sibling of catalogue qwen3-asr-1-7b. Proposed comparison-only model: evaluated by exact HF id on the Elm Arabic board and the Persian Triple Threat board, where it is the third model bridging the gap closer gemini-3-5-transcribe.

54.6±1.4#76 of 162 speech-to-text modelsfrom 23 benchmark results across 8 sources

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of Qwen3-ASR-0.6B, out of 176 with a score, then more from Alibaba (Qwen).

#	model	score	where	RAM	released
85	Parakeet CTC 1.1B NVIDIA	55±3	local	~5.6GB	2024-01
86	SeamlessM4T v2 Large Meta	55±4	local	~9.2GB	2023-11
87	Qwen3-ASR-0.6B Alibaba (Qwen)	55±1	local	~2.8GB	2026-01
88	Ink-Whisper Cartesia	55±3	api price unknown		2025-06
89	Breeze ASR 25 MediaTek Research	54±4	local	~1.7GB	2025-06
	more from Alibaba (Qwen)
39	Qwen3-ASR-1.7B Alibaba (Qwen)	60±1	api $0.0005/audio min +local	~3.3GB	2026-01

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

multilingual transcription▂··▃▄█▆▅▄▂·▂ #29 of 36 * · 27% of its score · 2 tests

Beats GPT-4o Mini Transcribe here, though they rank higher overall. Trails Omnilingual ASR CTC 1B v2, Omnilingual ASR CTC 7B v2, Omnilingual ASR CTC 3B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Open ASR multilingual WER	8.2% ✎	▂··▂▃▂█▃█▂▁▁ beats 5/29	avg 8.9%	▇▇▇▇▇▇▇▇▇ 23%	typical
QCRI Arabic-English ASR leaderboard, 22-set average WER	36.6% ✎	█···█······█ beats 0/2	avg 53%	▇▇ 4.0%	small run

other tests18 not in a task · 72% of its score

test	result	vs the others in its run	difficulty	counts	why
LibriSpeech test-clean WER +1	2.1% ✎	▁···▁▁▂·▄▄██ beats 23/56	avg 2.3%	▇▇▇▇ 10%	blunt · several runs
Open ASR English short-form WER (March 2026)	6.4% ✎	▁▁▁▃··▁▁▂██▆ beats 70/86	avg 8.2%	▇▇▇ 7.0%	blunt
RW-Voice-EQ ASR · 4 real-world English sets, mean WER	8.8% ✎	█▃▃▅▃···▃··▃ beats 5/9	avg 15%	▇▇ 5.5%	typical
FLEURS Mandarin test error rate +1	7.6% ✎	▂▂··▂▄▂▂·▆██ beats 16/17	avg 12%	▇▇ 5.0%	several runs
Persian ASR Triple Threat · VisualEars6669 + FLEURS-fa equal-weight WER	52.2% ✎	█·········██ beats 0/2	avg 43%	▇▇ 4.8%	sharp · noisy · small run
Multilingual LibriSpeech English test WER lab	6.0% ✎	█······█·█·█ beats 0/3	avg 7.7%	▇▇ 4.4%	typical
FLEURS English test WER +1	4.2% ✎	▁··▁▁▁▁▁▁▃█▁ beats 46/56	avg 5.8%	▇▇ 4.4%	blunt · several runs
WenetSpeech test_net error rate lab	6.0% ✎	█···█·····██ beats 2/3	avg 13%	▇▇ 4.1%	typical

10 more tests

test	result	vs the others in its run	difficulty	counts	why
AISHELL-2 test error rate lab	3.1% ✎	█··█·····█·█ beats 2/3	avg 5.7%	▇▇ 4.1%	typical
LibriSpeech test-other WER lab	4.5% ✎	▁▁▁▂▅▃▅▄█▂▂▁ beats 30/54	avg 5.5%	▇ 3.7%	typical
PESTE · Persian FLEURS (fa_ir) corpus CER	20.9% ✎	▄··▄·▄▄▄··██ beats 3/8	avg 12%	▇ 3.4%	typical
Common Voice English WER (version unspecified) lab	9.9% ✎	█··▄·······▄ beats 0/3	avg 13%	▇ 2.9%	typical
Common Voice Mandarin error rate (version unspecified) lab	6.9% ✎	█·······██·█ beats 1/3	avg 11%	▇ 2.8%	typical
GigaSpeech test WER lab	8.9% ✎	▁····▁··▁▄█▃ beats 16/17	avg 17%	▇ 2.8%	typical
WenetSpeech test_meeting error rate lab	6.9% ✎	█··█······██ beats 2/3	avg 21%	▇ 2.7%	sharp · noisy
VoxPopuli English test WER lab	10.0% ✎	█▄▄··████▄██ beats 1/16	avg 13%	▇ 2.2%	typical
Open Universal Arabic ASR Leaderboard · 6-test-set average WER	42.2% ✎	▃·▂▂▅█▅▃▂▅▂▂ beats 6/22	avg 40%	▇ 2.0%	blunt
TED-LIUM test WER (short-form) lab	3.9% ✎	▁·▁▇█▃·▁▁▃█▄ beats 22/27	avg 6.2%	▇ 0.2%	blunt

$ where to run it

Other ways to run it

Run it as published · needs ~2.76 GB, NVIDIA only · from huggingface.co/Qwen (official)
uvx --from 'qwen-asr[vllm]' qwen-asr-serve Qwen/Qwen3-ASR-0.6B --gpu-memory-utilization 0.8 --host 0.0.0.0 --port 8000

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3-ASR-1.7B model card 33%, Open ASR Leaderboard 30%, transcribe.cpp model catalog… 16% and 5 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Qwen/Qwen3-ASR-1.7B model card · lab-reported	13	▇▇▇▇▇▇▇▇▇▇ 33%	Apache-2.0
Open ASR Leaderboard	2	▇▇▇▇▇▇▇▇▇ 30%	Apache-2.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports	3	▇▇▇▇▇ 16%	MIT
RW-Voice-EQ Bench (Hume AI)	1	▇▇ 6%	CC BY 4.0
Persian ASR Triple Threat Leaderboard (Reza2kn)	1	▇ 5%	Apache-2.0
QCRI Arabic + English ASR leaderboard (Hugging Face Space QCRI/AURA-Arabic-English-ASR-leaderboard)	1	▇ 4%	CC BY 4.0
PESTE · Persian Speech to Text benchmark (fleurs-fa-ir-v1)	1	▇ 3%	Apache-2.0
Open Universal Arabic ASR Leaderboard (Elm Research Center)	1	▇ 2%	Apache-2.0
prior · pull toward the middle		▇ 1%
weighted average		= 54.6

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json