Qwen2-Audio-7B-Instruct: benchmark results
Alibaba's Qwen2-Audio-7B-Instruct, an instruction-tuned audio-language model supporting free-form voice chat and text-guided audio analysis from speech or sound inputs. Provider: Alibaba. Released 2024-07-31. Access: Open.
Unified ELO 1483 ± 10, rank #824 of 1639 rated models, from 192 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AOR-Bench | 17.8 | Over-refusal rate (%; refusals of 500 pseudo-harmful audio q | 100 |
| AOR-Bench - MB-Score | 82.8 | MB-Score (%; harmonic mean of the true-refusal rate on 500 h | 100 |
| Afrispeech Semantics - Consistency (AfriSpeech-200) | 84.46 | Macro-F1 (%; binary consistent versus inconsistent judgments | 100 |
| Afrispeech Semantics - Entailment (AfriSpeech-200) | 71.23 | Macro-F1 (%; three-way entailment, neutral or contradiction | 100 |
| Afrispeech Semantics - Entailment (AfriSpeech-Medical) | 55.34 | Macro-F1 (%; three-way entailment, neutral or contradiction | 100 |
| AudioBench - IMDA GR Sentence | 67.36 | Mean Score (%) | 100 |
| AudioBench - MuChoMusic | 71.61 | Score (%) | 100 |
| AudioBench Music | 65.89 | Average Score (%) | 100 |
| HELM Audio - Audiocaps | 2.67 | GPT-4o Judge Critique | 100 |
| HELM Audio - Knowledge | 90.62 | Mean win rate | 100 |
| HELM Audio - Air Bench Foundation | 72.39 | EM | 93.8 |
| HELM Audio - Air Bench Chat | 3.22 | GPT-4o Judge Critique | 92.9 |
Interactive version: theaggregate.ai/model?slug=qwen2-audio-7b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-09.