Stability AI spun out its audio research group on Sunday as Harmonic Labs, an independent company capitalized with $80 million from Coatue, Sound Ventures, and, notably, the venture arms of two major record labels. Stability retains a minority stake and licenses the existing Stable Audio models. The new company starts with 34 employees, most of the original research team, and something Stability spent three years failing to obtain: signed training data licenses with Universal Music Group and Warner Music Group covering a defined catalog for model training, with per-use royalty terms for commercial output.
Why the spinout was the price of the licenses
Stability's legal position made it impossible for a major label to sign anything. The company faces litigation from Getty Images in both the United States and the United Kingdom over image training data, and its public posture through 2023 and 2024 treated broad web scraping as defensible. Signing a training license with a company simultaneously arguing that licenses are unnecessary would have undermined the labels' position in every other negotiation they are conducting.
A new entity with clean provenance, no litigation history, and label investment solves that. Harmonic Labs will train on licensed catalog plus a corpus of production music it has separately acquired, and it will not use the scraped data that produced the original Stable Audio models. Chief executive Mira Voss, who led the Stability audio group, said the company will publish its data sources, which no major generative audio company currently does.
The economics of the licensing deal
Terms are private and two people familiar with the structure described the shape. Harmonic pays an upfront license fee and a royalty on commercial usage of generated output, with attribution flowing to rights holders whose material influenced a generation through a mechanism the company describes as contribution weighting. That mechanism is the technically interesting part and the part most likely to be contested, since attributing a generated waveform to specific training examples is not a solved problem.
Harmonic's approach uses influence estimation during training, maintaining a record of which catalog segments most affected the model's behavior in particular regions of its output space, and distributing royalties accordingly. It is approximate and the company says so. The labels apparently accepted an approximate mechanism with audit rights over an exact one that does not exist, which is a pragmatic decision and sets a template others will copy.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
What the product will be
Not a consumer music generator. Voss was explicit that the target is production tooling for people who already make music and media: stem separation, arrangement assistance, sound design for film and games, and generation of background music for advertising and corporate video where the alternative is a stock library. That segment has budget, has clearance requirements, and does not trigger the cultural objection that a system generating finished songs does.
The first product is a plugin for digital audio workstations, targeting Ableton, Logic, and Pro Tools, shipping in beta this autumn. Pricing has not been announced. A separate API aimed at game studios and advertising agencies follows in 2027. Voss said the company will not release model weights, a decision that departs sharply from Stability's history and that the licensing agreements almost certainly require.
The competitive field
Suno and Udio are the consumer-facing leaders and both spent 2024 and 2025 in litigation with the same labels now investing in Harmonic. Both eventually settled and signed licensing arrangements, which reshaped their products significantly and, according to people at both companies, constrained what they can generate. ElevenLabs has moved into music from voice with substantial resources. Adobe has a music generation product trained on licensed content, using the same approach it took with Firefly for images.
Adobe is the closest comparison and the most serious threat, because it has the distribution into creative professionals that Harmonic needs to build. Harmonic's counter is model quality, since the Stable Audio team produced genuinely strong results, and the argument that a specialist tool beats a feature inside a suite. That argument has succeeded and failed roughly equally often in creative software.
What this says about Stability
Stability has been shrinking and refocusing since its 2024 leadership change, selling or spinning out research groups and concentrating on image and video models for enterprise customers. The audio spinout continues that, and it converts an asset the company could not monetize into a stake in something that might work. That is a reasonable outcome for a company that spent 2023 as the most prominent name in open generative models and 2025 restructuring.
The broader lesson is about the cost of an early legal posture. Stability's original position attracted the researchers and the attention that made it briefly enormous, and it foreclosed the commercial relationships that would have made it durable. Companies founded since have watched that and have almost uniformly chosen licensing first, which is more expensive, slower, and available to fewer people.
The people affected most are researchers without institutional backing. A licensing-first world means training data costs money before a model exists, which favors companies that can raise capital against a plan and excludes anyone experimenting on a laptop. Several academics we spoke with described that as the quiet cost of the industry's legal settlement: the technical barrier to generative audio research fell to nearly zero over five years while the legal barrier rose to something only a funded company can clear. Harmonic Labs is a reasonable outcome for the researchers involved and not obviously a good one for the field.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- LLMs
