Suno launches v6, its first music model built with record-label support

Suno has released v6, its latest family of generative music models and the first one the company says was developed with support from the record industry. Jack Brody, a Suno executive, told The Verge that v6 was trained “from the ground up” with a new dataset that does not contain the same material used for earlier models. Suno says that dataset includes licensed content from Warner Music Group, BMG, and Believe, as well as user data.

That announcement is significant, but it does not resolve every question around training data. The available description does not establish that all material used for v6 is free of disputed or dubiously obtained content. The scope of the licenses, the treatment of user data, and the rights attached to generated music will remain important issues for artists and commercial users.

Three models with different goals

The v6 release is divided into three variants:

• v6 is the general-purpose model for higher-quality creation.
• v6-wild is intended to be less predictable and to produce more happy accidents and natural imperfections.
• v6-mini is the free, lower-resource option, optimized for speed rather than maximum detail.

In testing described by The Verge, all three models were substantially better at recognizing genre cues. Prompts involving styles such as hyperpop or krautrock, which had often missed the mark in older versions, produced tracks that captured more of the expected surface characteristics in v6. That improvement appears to be strongest in arrangement, timbre, and other recognizable genre signals. It should not be mistaken for a deeper understanding of musical history or intent.

The limits of manufactured imperfection

The most revealing weakness is that v6 still has trouble making music intentionally imperfect. Requests for off-key singing, loose rhythm, dissonance, monotone vocals, or a song without drums were often ignored or softened into a conventional, polished result. Even when asked to make a piano completely disregard key and rhythm, the model continued to produce music that was harmonically and rhythmically controlled.

The wild variant is designed to introduce more unpredictability, but the difference from standard v6 was not always obvious in limited testing. At the same time, v6 retains the kinds of imperfections associated with AI generation itself. Vocals can contain harsh-edged artifacts, and those artifacts appeared more noticeable in some cases than they were in v5.

This contrast highlights a broader challenge for music generators. They can reproduce the outward signals of a genre while remaining poor at controlling the small mistakes, timing shifts, and unstable intonation that make a human performance feel specific. A model may be able to imitate a style’s polish before it can convincingly recreate its flaws.

More than a prompt-to-song system

Suno is also changing how users interact with the product. In v6, people can use plain language in a chat interface to revise parts of a track. Adjusting a guitar line or changing a single lyric no longer necessarily requires regenerating the entire song. The model can also combine multiple elements from a user’s Suno library into new creations.

Prompts are no longer limited to text descriptions either. Users can start tracks from images, video, or other audio. These tools move Suno toward a more iterative editing workspace rather than a system that simply returns a finished song after one prompt. Questions remain about how reliably local edits preserve the rest of a composition and how licensed training material will affect future commercial use. V6 is rolling out gradually, and Suno says its older models will eventually be retired.

Source: The Verge AI