YuE2 and Suno, side by side.

What changes in your workflow?

We have not conducted a blind listening test of these services. Product features and plan terms change, so check Suno’s current offering before making a purchase decision.

Read the published scores with their conditions

The YuE2 authors report the following SongBench averages in their WildSongBench automatic evaluation dated 12 September 2026. Higher is better for this metric.

These are selected rows from the authors’ report, not our measurements. Standard YuE2 uses selection from two candidates; best-of-8 selects from eight. Both use the legacy benchmark decoder. One local generation with the listening decoder is not the same experiment.

A ranking is only part of the decision

The report covers 192 prompts and several metrics. Different metrics rank systems differently, and a small difference between averages does not establish statistical significance. Automatic scores also cannot tell you whether a particular voice or arrangement fits your song.

Compare complete songs in the languages and styles you actually use. Keep the prompt, number of attempts and selection rule consistent. Include failed and truncated outputs when judging how useful a workflow will be.

A practical way to choose

Try a hosted tool when you want to sketch a song without maintaining a runtime. Try YuE2 locally when you want to save intermediate artifacts, work with ABC scores and inspect the generation process.

Start with the listening examples, then check your hardware and the applicable license. Those three checks are more useful than a claim that one model wins every genre.