Project AURIONAI Use-Cases Research & Innovation
Alibaba (Qwen)
speech recognitionfamily

Qwen3-ASR

Alibaba Qwen multilingual ASR family (1.7B and 0.6B open, Flash via API). Most remaining language-ID errors on FLEURS come from confusing Malay and Indonesian.

Alibaba (Qwen)Alibaba (Qwen)China
Figure from the Qwen3-ASR paper: Qwen3-ASR family includes all-in-one ASR models with advantages in multilingual, noisy speech recognition, singing voice recognition and inference efficiency, s.
Levelfamily
Classspeech recognition
Parametersnot in the notes
Contextnot in the notes
Released29 Jan 2026
Opennessopen weights permissive
Runs onsingle datacentre GPU, CPU server
Statusactive

Benchmark Results

vendor-reportedmaintainer-run
FLEURSspeech
Qwen3-ASR-Flash-1208FLEURS Indonesian, Table A.2vendor-reported
3.65
Qwen3-ASR technical reportJan 2026 arxiv.org
Qwen3-ASR-1.7BFLEURS Indonesian, Table A.2vendor-reported
5.16
Qwen3-ASR technical reportJan 2026 arxiv.org
Qwen3-ASR-0.6BFLEURS Indonesian, Table A.2vendor-reported
7.92
048

Score in WER. Lower is better.

Qwen3-ASR technical reportJan 2026 arxiv.org
Hugging Face Open ASR Leaderboardspeech
Qwen3-ASR-1.7Beight-set macro-average English WERmaintainer-run
5.76
snapshot 2026-05-06 via benchmarklist6 May 2026 benchmarklist.com
Qwen3-ASR-0.6Beight-set macro-average English WERmaintainer-run
6.42
048

Score in WER. Lower is better.

snapshot 2026-05-06 via benchmarklist6 May 2026 benchmarklist.com
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).

Lineage

finetunecontinued pretrain
Qwen3 to Qwen3-ASR: finetune, low confidenceQwen3 to Qwen3-Embedding-0.6B: finetune, medium confidenceQwen3 to Qwen3-Embedding-4B: finetune, medium confidenceQwen3 to Qwen3-Embedding-8B: finetune, medium confidenceQwen3 to Qwen3-Reranker-0.6B: finetune, medium confidenceQwen3 to Qwen3-Reranker-4B: finetune, medium confidenceQwen3 to Qwen3-Reranker-8B: finetune, medium confidenceQwen3 to Qwen3-VL-Embedding-2B: finetune, medium confidenceQwen3 to Qwen3-VL-Embedding-8B: finetune, medium confidenceQwen3 to Qwen3-Swallow: continued pretrain, medium confidenceQwen3Qwen3Alibaba (Qwen), Apr 2025Qwen3-Embedding-0.6BQwen3-Embedding-0.6BAlibaba (Qwen), Jun 2025Qwen3-Embedding-4BQwen3-Embedding-4BAlibaba (Qwen), Jun 2025Qwen3-Embedding-8BQwen3-Embedding-8BAlibaba (Qwen), Jun 2025Qwen3-Reranker-0.6BQwen3-Reranker-0.6BAlibaba (Qwen), Jun 2025Qwen3-Reranker-4BQwen3-Reranker-4BAlibaba (Qwen), Jun 2025Qwen3-Reranker-8BQwen3-Reranker-8BAlibaba (Qwen), Jun 2025Qwen3-SwallowQwen3-SwallowInstitute of Science Tokyo, 2026Qwen3-ASRQwen3-ASRAlibaba (Qwen), Jan 2026Qwen3-VL-Embedding-2BQwen3-VL-Embedding-2BAlibaba (Qwen), Apr 2026Qwen3-VL-Embedding-8BQwen3-VL-Embedding-8BAlibaba (Qwen), Apr 2026
20262027Qwen3-ASR-0.6BQwen3-ASR-0.6B, Jan 2026Qwen3-ASR-1.7B, Jan 2026

1 member has no release date in the notes, so it is not on this timeline.

Alibaba (Qwen)Qwen3 is the parent (finetune)low confidencenotes
Alibaba (Qwen)Qwen3-ASR-0.6B belongs to this familyspeech recognition29 Jan 2026600M
Alibaba (Qwen)Qwen3-ASR-1.7B belongs to this familyspeech recognition29 Jan 20261.7B
Alibaba (Qwen)Qwen3-ASR-Flash-1208 belongs to this familyspeech recognition
Where this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 11 models. Confidence and evidence are the notes' own.

R&D Topics

Audio language models for speech understandingexemplar

Identity

Developer
Alibaba (Qwen)
Family
not in the notes
Released
29 Jan 2026 sourcesannouncement page
Weights published
–
Record checked
10 Oct 2026
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.

Architecture and Size

Architecture
autoregressive Audio encoder + Qwen3 decoder (unverified detail)
Total parameters
– sources
Active parameters
–
Checkpoint parameters
–
Context length
not in the notes
Input
audio
Output
text
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).

Openness and Access

Openness
open weights permissive
Licence
Apache License 2.0 sources
Access routes
open weights download
Where this comes from
The licence file governs over registry tags and aggregator labels.

Deployment

Runs on
single datacentre GPU, CPU server
Footprint
not in the notes
Runtimes
not in the notes
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.

Indonesia

Indonesian support
measured on a benchmark
Regional languages
not in the notes
Jakarta hosting
unknown
Languages
52 languages and dialects per Gladia; technical report: 30 languages and 22 Chinese dialects, Indonesian included
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.

Industry Use Cases

No industry use case links this model.

Sources

  1. Gladia - Best open-source speech-to-text models in 2026 gladia.io https://www.gladia.io/blog/best-open-source-speech-to-text-models
  2. Open ASR Leaderboard Benchmark Scores & AI Model Leaderboard | BenchmarkList benchmarklist.com https://benchmarklist.com/benchmarks/open_asr_leaderboard/
  3. Qwen3-ASR Technical Report arxiv.org https://arxiv.org/html/2601.21337v2

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Qwen3-ASR", https://aurion.semai.id/models/alibaba-qwen3-asr/, records of October 2026.