Qwen3-ASR

Alibaba's Apache 2.0 speech-to-text pair (1.7B, 0.6B): 30 languages plus 22 Chinese dialects, singing, streaming and timestamps.

In short. Qwen3-ASR turns speech into written text. It understands 30 languages, 22 regional Chinese dialects and many English accents, and it can even transcribe singing and songs with background music. It comes in a larger, more accurate size and a smaller, much faster one, and the Apache 2.0 license lets you use it commercially.

Summary

Alibaba's Qwen team released the Qwen3-ASR family on 29 January 2026 under the Apache 2.0 license. It has two speech recognition models, Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, plus a separate Qwen3-ForcedAligner-0.6B that predicts timestamps. According to the model card, both ASR models identify the language and transcribe 30 languages and 22 Chinese dialects, plus English accents from many countries, and they handle speech, singing voice and songs with background music. A single model runs in both streaming and offline modes and can transcribe long audio.

Both ASR models build on the audio understanding of Qwen's Qwen3-Omni foundation model. Qwen says the 1.7B reaches state-of-the-art accuracy among open-source recognizers and is competitive with the strongest commercial APIs, while the 0.6B trades some accuracy for speed and reaches about 2,000 times real-time throughput at a concurrency of 128; both claims come from Qwen's own benchmarks and technical report. The forced aligner is a non-autoregressive model that predicts timestamps for arbitrary units inside up to five minutes of speech in 11 languages.

The 1.7B checkpoint has 2.35 billion parameters in total and is a 4.7 GB BF16 download, the 0.6B has 0.94 billion and about 1.9 GB, and the aligner about 1.8 GB. Qwen provides a qwen-asr Python package with a Transformers backend and a vLLM backend for batch inference, asynchronous serving and streaming, plus an official Docker image. Since 26 June 2026 there are also -hf checkpoints with native Transformers support.

Good for

  • Transcribing meetings, calls, interviews and videos in Chinese, English and 28 other languages, including regional Chinese dialects
  • Live captioning and voice interfaces, using streaming mode
  • High-volume batch transcription with the faster 0.6B model on a modest GPU
  • Subtitles and karaoke-style alignment with word or character timestamps from the forced aligner

Watch out for

  • The best-in-class and throughput claims are Qwen's own; test the models on your own audio and accents
  • Coverage is 30 languages plus 22 Chinese dialects, so check that your language is on the list before you commit
  • The model card does not describe speaker labeling, so use a separate diarization tool if you need to tell speakers apart
  • Recognizers can still mishear names and numbers, so review transcripts that matter

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Qwen3-ASR. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
ForcedAligner 0.6B (timestamps) BF16 918M 1.84 GB 2.12 GB Laptop or phone
0.6B BF16 938M 1.88 GB 2.16 GB Laptop or phone
1.7B BF16 2.35B 4.7 GB 5.4 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Qwen3-ASR?
Qwen3-ASR turns speech into written text. It understands 30 languages, 22 regional Chinese dialects and many English accents, and it can even transcribe singing and songs with background music. It comes in a larger, more accurate size and a smaller, much faster one, and the Apache 2.0 license lets you use it commercially.
Can I use Qwen3-ASR commercially?
Yes. Qwen3-ASR is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Qwen3-ASR open source?
Qwen3-ASR is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Qwen3-ASR on my own computer?
Yes. The smallest variant, ForcedAligner 0.6B (timestamps), needs about 2.12 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Qwen3-ASR released?
Qwen3-ASR was released by Alibaba Qwen on January 29, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone
Parakeet TDT 0.6B v3 NVIDIA 627M – CC BY 4.0 (Permissive) 2.89 GB Laptop or phone
Cohere Transcribe Cohere 2.07B – Apache 2.0 (Permissive) 4.75 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU