Qwen3 Omni 30B A3B Instruct: benchmark results
Multimodal Instruct checkpoint containing both Thinker and Talker, with audio, video, image and text input and speech/text output. The technical report lists a 30B-total/3B-active Thinker and a separate 3B-total/0.3B-active Talker, plus audio/vision encoders, MTP and waveform modules. The 3B active label describes the Thinker, not all modalities and output paths. Provider: Alibaba. Released 2025-09-22. Access: Open.
Aggregate Rating 1621 ± 7, rank #383 of 1656 rated models, from 299 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BASIS-Bench - ASR Fair-Speech Accent | 95.3 | 100-WER | 100 |
| BASIS-Bench - ASR Fair-Speech Gender | 94.7 | 100-WER | 100 |
| BASIS-Bench - ASR SD-QA Accent | 95.9 | 100-WER | 100 |
| BASIS-Bench - SQA LibriSQA Gender | 83.5 | Accuracy (%) | 100 |
| BASIS-Bench - SQA SD-QA Accent | 81.6 | Accuracy (%) | 100 |
| EgoIntrospect - Memory Recall Prediction | 55.5 | Accuracy (%) on 200 four-choice questions asking which detai | 100 |
| KoALa-Bench - PA-QA | 92.51 | Accuracy (%) on PA-QA, MCTest stories read as long Korean sp | 100 |
| MCGA | 442.9 | Score (0-600; sum of the six task scores on 1,000 test recor | 100 |
| MCGA - ASR - Ci | 2.8 | Character error rate (%; 200 recorded ci lyrics of the test | 100 |
| MCGA - ASR - Qu | 4.1 | Character error rate (%; 200 recorded qu verses of the test | 100 |
| MCGA - Speech Emotion Captioning | 58.4 | Emotion Caption Fidelity (0-100; persona, global tone and se | 100 |
| MCGA - Speech Understanding | 86.9 | Accuracy (%; multiple-choice questions on the content of 1,0 | 100 |
Interactive version: theaggregate.ai/model?slug=qwen3-omni-30b-a3b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.