Qwen3-ASR vs Voxtral Realtime: which streaming model should you use?
Qwen3-ASR covers 30 languages plus 22 Chinese dialects with timestamps; Voxtral Realtime is one 4B model with an 80 ms to 2.4 s delay. Compare languages and latency.
Verdict. Both are Apache 2.0 speech-to-text models with streaming, so choose by languages and by how you serve them. Qwen3-ASR comes in 1.7B and 0.6B sizes, covers 30 languages plus 22 Chinese dialects, transcribes singing and has a forced aligner for word timestamps. Voxtral Mini Realtime is one 4B model for 13 languages in which you set the delay between 80 ms and 2.4 s: Mistral's table shows an average error rate of 12.60% at 160 ms and 8.72% at 480 ms. Neither model card describes speaker labels, and both developers' accuracy claims are their own.
Specifications
| Specifications | Qwen3-ASR | Voxtral Realtime |
|---|---|---|
| Developer | Alibaba Qwen | Mistral AI |
| Released | January 29, 2026 | February 4, 2026 |
| Total parameters | 2.35B (1.7B) | 4.43B (Voxtral Mini 4B Realtime) |
| Active parameters | All | All |
| Architecture | Other | Dense |
| Context window | – | 128K tokens |
| Input | Audio | Audio |
| License | Apache 2.0 (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 2.12 GB (ForcedAligner 0.6B (timestamps)) | 10 GB (Voxtral Mini 4B Realtime) |
Performance claims come from the developer and have not been independently verified.
Pick Qwen3-ASR if
- Your audio is Chinese, includes regional dialects or singing.
- You need word or character timestamps for subtitles.
- You want a small, fast 0.6B model for high-volume batch jobs.
- You need a language outside Voxtral's 13.
Pick Voxtral Realtime if
- You need live captions or a voice assistant and want to set the delay yourself.
- You want one 4B model for private or on-device setups under Apache 2.0.
- The 13 widely spoken languages cover your needs.
- You can run vLLM with websocket streaming, which the model card recommends.