MusTBENCH - Temporal Source Grounding: leaderboard

MusTBench (2026) tests temporal grounding of audio-language models in full-length music tracks, zero-shot. Temporal source grounding: give the time at which a named instrument or the vocals first start or last stop. Mean of onset and offset hit rates within 3 seconds, in percent; higher is better.

Metric: Temporal Source Grounding (TSG, 400 questions): the model gives the time at which a named instrument or the vocals first start or last stop; score is the mean of onset and offset Hit@3s, the share of predicted timestamps within 3 seconds of the ground truth (%, 0-100), on the MusTBench test set (full-length MTG-Jamendo and Slakh2100 tracks, music-expert-validated questions), zero-shot with the full track and a task instruction, temperature 0, unparseable timestamps and intervals counted wrong; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Flash65.8#463
2Gemini 2.5 Pro57.3#276
3Gemini 3 Pro52.3#145
4Gemini 3 Flash48.8#175
5Music Flamingo38.5#866
6Qwen 3 Omni 30B A3B36#847
7Qwen2.5-Omni-7B21.3#1884
8GPT-Audio11.5#607
9AF-Next10
10GPT-Audio 1.58.8#501
11Phi-4 Multimodal Instruct7.3#4154
12Qwen2.5-Omni-3B4.8#3519

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mustbench-temporal-source-grounding · How It Works · Data refreshed daily, snapshot 2026-11-21.