What is Qwen3-ASR (1.7B and 0.6B)?
Qwen3-ASR is a pair of open speech recognition models from Alibaba's Qwen team, posted on Hugging Face on 28 January 2026 under Apache 2.0: Qwen3-ASR-1.7B and the smaller Qwen3-ASR-0.6B. Both are built on the audio side of Qwen3-Omni and handle 30 languages plus 22 Chinese dialects, identify the language themselves, and run either offline or streaming from the same checkpoint. Qwen says they also transcribe singing and songs with background music.
Timestamps come from a separate Qwen3-ForcedAligner-0.6B, which aligns words or characters in up to 5 minutes of speech in 11 languages. The qwen-asr Python package wraps both, with a Transformers backend and a faster vLLM backend, and Qwen ships a Docker image and Gradio demos. The 1.7B recorded about 1.7 million Hugging Face downloads in the past month and the 0.6B about 714,000.
Is Qwen3-ASR (1.7B and 0.6B) worth it?
Qwen3-ASR-1.7B is the most capable general-purpose open speech model for most teams, because it pairs leaderboard-level English with coverage nobody else in open weights matches. In Qwen's tables it scores 1.63% and 3.38% on LibriSpeech clean and other, against 1.51% and 3.97% for Whisper large-v3, and 4.97% on WenetSpeech net, where GPT-4o-Transcribe scored 15.30%. On the KeSpeech dialect set it reaches 5.10% against Whisper's 28.79%. These are Qwen's measurements with no language hint given.
Outside Qwen's own numbers the picture holds. On the Open ASR Leaderboard Cohere's March launch table put the 1.7B at 5.76% average English WER, fifth behind Cohere Transcribe; by October IBM's chart of the same leaderboard showed the Transformers-native Qwen3-ASR-1.7B-hf first among open-weight models at 4.94%, ahead of Granite Speech 5.0 at 5.06%. Artificial Analysis does not list the open model, so there is no independent like-for-like test with commercial APIs.
The 0.6B is the throughput option. Qwen reports 2,000 times real-time throughput at a concurrency of 128, and it trails the 1.7B by up to 2.5 points on English sets (9.92% against 7.39% on Common Voice English) and on dialects (7.08% against 5.10% on KeSpeech). Hosted on OpenRouter, the 1.7B costs about $0.45 per 1,000 minutes and the 0.6B about $0.20, both well under Whisper large-v3-turbo on Groq at $0.67.
The catch: features are spread across modes. Streaming runs only on the vLLM backend and then supports neither batching nor timestamps; timestamps need the second aligner model and stop at 5 minutes of audio per call. Long files need a larger max_new_tokens or the transcript is cut short, and the documented demo and serving commands bind to 0.0.0.0 on port 8000, so put them behind a firewall.
Our take: for anything involving Mandarin, Cantonese or regional Chinese, Qwen3-ASR is the first model to try; for English alone, it is now level with the best open transcribers rather than clearly ahead of them.
Where it earns its place
A data team at a Shenzhen customer-service outsourcer
Transcribe calls in Mandarin, Cantonese and Sichuanese for quality review
Run the 1.7B on vLLM with language left to auto-detect; Qwen reports 3.98% on FLEURS Cantonese and 11.99% on easy Sichuanese. Self-hosting keeps call audio in-house, and Apache 2.0 needs no licence negotiation.
A developer at a karaoke and lyrics app
Generate time-coded lyrics from songs with backing tracks
Transcribe with the 1.7B, which Qwen trained on singing and songs with background music, then pass each verse through Qwen3-ForcedAligner-0.6B for word timings, in chunks of up to 5 minutes.
A startup building a voice agent for Southeast Asian markets
Live transcription in Thai, Vietnamese, Indonesian and Malay
Serve the 0.6B in streaming mode on vLLM for partial results; hosted, the same model costs about $0.20 per 1,000 minutes on OpenRouter. Streaming disables timestamps, so turn-taking has to rely on a separate voice-activity detector.
Before you commit
- Setup
- pip install -U qwen-asr (add the vllm extra for speed and streaming), then Qwen3ASRModel.from_pretrained('Qwen/Qwen3-ASR-1.7B') and transcribe; FlashAttention 2 is recommended. vllm serve Qwen/Qwen3-ASR-1.7B also works.
- Integrations
- qwen-asr Python package, vLLM with OpenAI chat and transcription APIs, Docker image qwenllm/qwen3-asr, Gradio demos; hosted on OpenRouter and Alibaba Cloud Model Studio, whose qwen3-asr-flash takes files up to 10 MB and 5 minutes.
- Data and privacy
- Self-hosted under Apache 2.0, audio stays on your hardware. The documented demo and server commands listen on 0.0.0.0:8000, so they are reachable on your network unless firewalled; the browser demo needs HTTPS for microphone access.
- Cost at scale
- Self-hosted the 0.6B fits a small GPU and the 1.7B a mid-range one. Hosted on OpenRouter, a million minutes a month is about $450 on the 1.7B or $200 on the 0.6B.
- Limits
- No diarization; timestamps only via the aligner, up to 5 minutes and 11 languages; streaming is vLLM-only with no batching or timestamps; long audio needs max_new_tokens raised.
What’s good
- 30 languages and 22 Chinese dialects with automatic language ID
- Apache 2.0 for both sizes
- 5.10% on KeSpeech dialects against 28.79% for Whisper large-v3 in Qwen's tests
- Streaming and offline from one checkpoint; 0.6B reaches 2,000 times throughput at concurrency 128
- About $0.45 per 1,000 minutes hosted
What’s not
- Timestamps need a separate aligner, capped at 5 minutes and 11 languages
- Streaming mode cannot batch or return timestamps
- Most benchmarks are Qwen's own; not on Artificial Analysis
- No speaker labels
- Documented serving commands listen on 0.0.0.0
Best for
Developers who need one open model for multilingual and Chinese-dialect transcription, batch or live.
Qwen3-ASR (1.7B and 0.6B) alternatives
Qwen3-ASR (1.7B and 0.6B): common questions
- Is Qwen3-ASR (1.7B and 0.6B) worth it?
- Qwen3-ASR is the open transcriber with the widest reach, and the one to beat for Chinese in any of its forms. We rate it 4.3 out of 5.
- How much does Qwen3-ASR (1.7B and 0.6B) cost?
- Free under Apache 2.0. Hosted on OpenRouter (DeepInfra, Prime Intellect): 1.7B at $0.0000075 a second of audio (about $0.45 per 1,000 minutes), 0.6B at about $0.20 per 1,000 minutes.
- What are the downsides of Qwen3-ASR (1.7B and 0.6B)?
- Timestamps need a separate aligner, capped at 5 minutes and 11 languages; streaming mode cannot batch or return timestamps; most benchmarks are Qwen's own; not on Artificial Analysis; no speaker labels; documented serving commands listen on 0.0.0.0.
- Who is Qwen3-ASR (1.7B and 0.6B) best for?
- Developers who need one open model for multilingual and Chinese-dialect transcription, batch or live.
- How much does Qwen3-ASR (1.7B and 0.6B) cost for a team or in production?
- Self-hosted the 0.6B fits a small GPU and the 1.7B a mid-range one. Hosted on OpenRouter, a million minutes a month is about $450 on the 1.7B or $200 on the 0.6B.
- Is Qwen3-ASR (1.7B and 0.6B) safe for company data?
- Self-hosted under Apache 2.0, audio stays on your hardware. The documented demo and server commands listen on 0.0.0.0:8000, so they are reachable on your network unless firewalled; the browser demo needs HTTPS for microphone access.
- What are the best Qwen3-ASR (1.7B and 0.6B) alternatives?
- The closest alternatives we have reviewed are Cohere Transcribe (03-2026), Granite Speech 5.0 2B, Whisper large-v3-turbo, VibeVoice-ASR and Nemotron 3.5 ASR Streaming 0.6B.
Answers checked Nov 7, 2026.
Make Qwen3-ASR (1.7B and 0.6B)? Show this rating on your site
