Qwen2-Audio Technical Report

Audio-language models needed complex hierarchical task tags, which made pre-training data curation brittle and instruction-following weak.

Replaces hierarchical tags with NATURAL LANGUAGE PROMPTS describing each data type and task, unifies voice-chat and audio-analysis interaction modes without system-prompt switching, and applies DPO to improve factuality and instruction adherence.

Reports outperforming prior SOTA including Gemini-1.5-pro on audio-centric instruction-following subsets of AIR-Bench. This is a model-quality claim with no latency, memory or hardware disclosure, so no serving speedup can be extracted.

Technical report, not a peer-reviewed paper; no venue found. Qwen2-Audio weights are widely downloaded and fine-tuned, and Qwen2-Audio-family inference exists in production stacks, but the model's serving cost is dominated by the audio encoder's sequence length, which this report does not characterise. Note the arXiv list shows a much later Qwen-Audio 3.x line (2607.11738 Qwen-Audio-VAE, 2609.07549 Qwen-Audio-3.0-ASR, 2609.25176 Realtime) - those are separate technical reports outside this record. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.