Tier 3 Practical open-weight
What the publisher claims: open-weight, open-source
The one-liner: Apache-2.0 speech recognition with a 0.6B model small enough for on-device use.
The six questions
Both checkpoints are downloadable from the publisher's Hugging Face repositories.
Apache-2.0, an OSI-approved permissive license with no revenue or use threshold.
Transformers-native and small enough for local or edge deployment, including a 0.6B checkpoint that fits comfortably on-device.
Model cards report leaderboard results with mean word error rates of 5.59% for the 1.7B model and 6.31% for the 0.6B, plus supported-language coverage; training data and recipe are not published.
Limited.
The Qwen3-ASR release ships weights and cards on Hugging Face with essentially no public development repo (the QwenLM/Qwen3-ASR GitHub repo does not exist), so outside input is limited to Hub discussions that Alibaba reads.
License
Apache-2.0 OSI-approved — license source ↗
- Standard Apache-2.0 obligations only
What is actually released
- weights (0.6B, 1.7B)
- forced aligner for word-level timestamps
- model cards
- evaluation results
Why it matters
Shows speech recognition quietly becoming a permissively licensed commodity, which changes the strategic position of commercial transcription APIs.
Sources
- https://huggingface.co/Qwen/Qwen3-ASR-1.7B
- https://huggingface.co/Qwen/Qwen3-ASR-0.6B
- https://benchlm.ai/models/qwen3-asr-1-7b
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.