A self-hosted model and an API service trade cost for licence scope

The audio task

A small studio adding a generative or assistive AI tool to its stack has two structurally different routes: run an open-source model on its own hardware, or call a commercial API and pay per use. The choice affects more than convenience. It changes who controls the compute, what happens to submitted data, and, critically, what a studio is actually licensed to do with whatever the model outputs.

What the documents show

Meta's audiocraft repository hosts MusicGen, installable locally with a documented pip-based setup; its own model card states the code is released under the MIT licence while 'the model weights are released under CC-BY-NC 4.0,' and names the model's primary intended use as research rather than client production work. Google's Gemini API pricing page lists its Lyria 3.5 music model at a flat '$0.08 per song' with no free tier for that model, and states in its own disclosure column that content submitted under a free tier is used to improve Google's products while paid-tier use is not.

Rights status

These are two different rights postures, not just two prices. MusicGen's weights carry a noncommercial licence on the model itself, according to its own model card; a studio self-hosting MusicGen for a paying client's release should treat that restriction as applying to the model, separate from whatever rights exist in the studio's own generated output. Google's Lyria pricing page does not state an output-rights grant on the page itself; it addresses billing and data usage, not licence to the generated audio, which would need confirming in Google's separate terms of service.

What to check before you use it

This is an editorial comparison built from named projects' and vendors' own documentation, not a benchmark of output quality. Before choosing a route, check the model's or service's own stated licence for both access and output; confirm whether self-hosting requires hardware the studio does not have; and check whether an API's paid tier, not just its free tier, changes what happens to submitted material, since the two can differ on the same page.

• Does the model's own licence permit the specific commercial use this project needs?
• Is the studio equipped to run and maintain a self-hosted model, or does that shift cost into staff time instead of a subscription?
• Does the API's paid tier state, in its own terms, that submitted audio or prompts are not used to train future models?

Neither route is inherently cheaper or safer; each vendor's or project's own documentation states a different trade between control, cost, and licence scope, and that documentation is what to compare, not a marketing claim about ease of use.