YuE2 48 kHz stereo on a 24GB GPU vs a hosted Sume music job

The YuE2-3B card says it runs 48 kHz stereo songs on a 24GB GPU without quantization, and its quick start asks for Linux, Python 3.10+ and a 24GB NVIDIA GPU with BF16 support. A hosted Sume music job needs none of that: you send a prompt of 1-5000 characters and read the audio artifact when the job finishes.

Hardware facts are from the model card, read 2026-10-01. Hosted facts are from Music Router, read 2026-10-01.

What does the local setup need?

The card's quick start installs yue2_infer from a wheel, then loads YuE2Pipeline onto CUDA. A call takes style, lyrics and a seed, and the result saves as a FLAC file.

How does a hosted job finish?

Create the job with POST /v1/music-router/generate, then poll GET /v1/jobs/{id}/status and read GET /v1/jobs/{id}/result. The audio is the artifact in result.artifacts[] where type is audio. The docs show an Idempotency-Key header on the create call.

What makes a retry safe?

The Music Router docs show an Idempotency-Key header on the create call. The video docs say that header makes retries safe and that a replay returns the original job. Whether music replays behave identically is not stated on the music page, so confirm it in the API reference before relying on it. See idempotency keys for AI video apis for the pattern.

When is local the better fit?

When you already own the GPU, want the editable-score workflow, and your use is non-commercial. When you want a job with no hardware and a fixed price per generation, use the hosted route; the docs say every Music Router model charges the fixed Music price per audio generation. See also open weights versus a hosted music API.