- New local ASR models: Qwen3-ASR 0.6B and 1.7B (WebGPU) (#467, #469, #473). 16 languages including Chinese, English, Japanese, Korean and Cantonese. Downloads 0.9 GB / 2.0 GB (fp16 GPUs) or 1.3 GB / 2.7 GB. The 1.7B roughly halves Japanese errors at ~1.3× the decode time. WebGPU required. Models: 0.6B, 1.7B.
- Recommended models (#471). Both Qwen3-ASR sizes are now recommended; Voxtral Mini 3B 2507 is not (still available). Thai, Indonesian and Cantonese default to Qwen3-ASR; other defaults and your existing selections are unchanged.
Full Changelog: v0.39.3...v0.40.0