Published
Updated
Qwen-Audio-Jev prints 78.70% on MMAU-mini and about 20x the base model's speed
Qwen-Audio-Jev prints 78.70% on MMAU-mini and a 20x speedup against Qwen-Audio-3.1-ASR-Flash-Next. The paper is arXiv 2610.21614, submitted October 15, 2026. This desk did not run the model.
1LittleCoder posted Qwen-Audio-Jev on October 18, 2026, at 21:14 UTC. The text quotes the abstract: a dedicated Jev decision head on Qwen-Audio-3.1-ASR-Flash-Next, about 20 times faster than that base model, Choice 78.70% on MMAU-mini and 84.06% on MMSU, Noul equal error rate 6.50% on VoxCeleb1-O, Score normalized MAE 0.1062 on NISQA overall with Pearson 0.8210 and Spearman 0.8074. The chart on that post is the image saved for this page.
The paper is arXiv 2610.21614, submitted 15 October 2026. Authors Ying Fang, Zhifu Gao, Qing Gu, Jiajun He, Kangqi Jing, Xiangang Li, Yu Pu, Wupeng Wang. The abs page classifies it under Computer Science, Sound.
The model reads audio, a natural-language question, and an explicit answer specification. It returns a distribution or a score. The three primitives are Choice, Noul, and Score. The HTML we opened prints eighteen labeled configurations. The Choice table also lists CochlScene 77.87, NSynth instrument 72.29, NonSpeech7k 92.69, DiDiSpeech gender 100.00, MELD emotion 57.31, and MMAR 69.90. We found no public weight file on that abs page. We did not run the model.
ToddlerJev reprints 78.70, 84.06, and 69.90 as published Qwen-Audio-Jev cells. Its own scores on those three names are 68.70, 61.30, and 56.90. That blog says the comparator values are published results, not reruns.
AudioJev, arXiv 2610.01293, submitted 1 October 2026, is a third paper. It prints 68.88% and 55.33% mean accuracy on complete MMAU and MMAR across three training seeds.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
1LittleCoder, @1littlecoder, October 18, 2026, 21:14 UTC, status 2111928938922647904. The post quotes the abstract: about 20x inference speedup against Qwen-Audio-3.1-ASR-Flash-Next, Choice 78.70% on MMAU-mini and 84.06% on MMSU, Noul equal error rate 6.50% on VoxCeleb1-O, Score normalized MAE 0.1062 on NISQA overall with Pearson 0.8210 and Spearman 0.8074. The share image on this page is the chart from that post. The paper link in the post is truncated. The abs page we opened is https://arxiv.org/abs/2610.21614.
arXiv:2610.21614, submitted 15 October 2026, Computer Science > Sound. Authors Ying Fang, Zhifu Gao, Qing Gu, Jiajun He, Kangqi Jing, Xiangang Li, Yu Pu, Wupeng Wang. Token Foundry, Alibaba Group on the alphaXiv listing. The abstract matches those four figures and the 20x line. The HTML table for Choice also prints CochlScene 77.8717, NSynth instrument 72.2900, NonSpeech7k 92.6897, DiDiSpeech gender 100.0000, MELD emotion 57.3072, and MMAR 69.9000. Eighteen labeled configurations. The model adds a Jev decision head on Qwen-Audio-3.1-ASR-Flash-Next. Primitives Choice, Noul, Score. We found no public weight file on that abs page. We did not run the model.
Compare
78.70% is the paper's MMAU-mini Choice accuracy. ToddlerJev's blog reprints 78.70, 84.06, and 69.90 as published Qwen-Audio-Jev cells and scores itself 68.70, 61.30, and 56.90 on those three names. The ToddlerJev page says those comparator values are published results, not reruns, and that protocols differ.
AudioJev, arXiv 2610.01293, submitted 1 October 2026, prints 68.88% and 55.33% mean accuracy on complete MMAU and MMAR across three seeds. That is a different paper.
6.50% EER on VoxCeleb1-O is the paper's Noul speaker-verification cell. A reply on the 1LittleCoder thread said ECAPA-TDNN is about 1%. We did not open that baseline.
20x is the paper's speedup against the original autoregressive Qwen-Audio-3.1-ASR-Flash-Next, not a clock against TypeSafe Jev.
Terms
- Qwen-Audio-Jev
- A Jev decision head on Qwen-Audio-3.1-ASR-Flash-Next. arXiv 2610.21614, submitted October 15, 2026. Choice, Noul, and Score over audio. The abstract prints 78.70% on MMAU-mini and about 20x the base model's speed.
- MMAU-mini
- A multiple-choice audio understanding set. Qwen-Audio-Jev prints 78.70%. ToddlerJev prints 68.70% on 1,000 items. Those two cells are not one run.
- VoxCeleb1-O EER
- Equal error rate on a speaker-verification split. Qwen-Audio-Jev prints 6.50% for its Noul setup.