Google: Lyria 3 Clip Preview

Google: Lyria 3 Clip Preview is a music generation model from Google, part of the Lyria 3 family. It is made available through the Gemini API, and Google describes Lyria 3 as its family of music generation models. The model sits in the audio category and also carries text capabilities, so it behaves as a multimodal generator rather than a pure text model. If you are building tooling that needs to produce short musical clips, this is the Google offering aimed at that job.

The model accepts text and image inputs and can return both text and audio. That means a prompt describing a style, mood or instrumentation can be paired with a reference image, and the output can include generated audio alongside text. The context window is 1,048,576 tokens, which is large enough to hold lengthy prompts, detailed descriptions or supporting material in one request. The maximum output is 65,536 tokens. This model is explicitly a clip generator: Google's own description refers to 30 second duration clips, so expect output in that short form rather than full length tracks.

Practical use cases follow from the clip length and the image input. You can generate 30 second background beds for video, podcast intros and outros, game or app ambience, or social media content. Because an image can be supplied, you can steer generation from cover art, a mood board or a storyboard frame, which suits workflows where the visual direction already exists. The text output can carry descriptive or supporting content back to your pipeline. Teams that need a quick musical idea to iterate on, rather than a finished production master, are the natural fit.

Calling it through api14 works the same way as any other OpenAI-compatible endpoint. Use the model id "google/lyria-3-clip-preview" in your request. The supported parameters are max_tokens, response_format, seed, temperature and top_p. Use seed when you want reproducible generations, temperature and top_p to control variation, response_format to shape the returned output, and max_tokens to bound generation. Because the endpoint is OpenAI-compatible, existing SDKs and client code can point at api14 without a rewrite.

One constraint to keep in mind is duration. The model generates 30 second clips, so if your project needs longer continuous music you will need to stitch multiple generations or look at a different option. It is also a preview release from Google, which is worth weighing if you need long term stability guarantees. For short, prompt driven or image guided music clips, it is a direct fit.