tongflow-modal-levo

July 4, 2026 · View on GitHub

Official TongFlow plugin. Text/lyrics-to-music generation with LeVo 2 / SongGeneration (lglg666/SongGeneration-v2-large), running on a GPU via Modal.

LeVo 2 is a commercial-grade, multilingual (zh, en, es, ja, …) song-generation model. It turns structured lyrics plus style tags into a full song with vocals and accompaniment.

Capabilities

  • Music generation (gen-music) — generate a song from lyrics + style tags.

Input tips

  • Lyrics use LeVo's structured format: sections separated by ;, e.g. [intro-short] ; [verse] First line. Second line. ; [chorus] Hook line. ; [outro-short]. [verse]/[chorus]/[bridge] carry lyrics; [intro-*]/[inst-*]/[outro-*] are instrumental.
  • Style tags go in the tags field as comma-separated keywords (gender, genre, emotion, instrument), e.g. female, pop, warm, piano. Avoid full sentences.

Credentials

Add in TongFlow Settings (gear icon, top-right):

KeyRequiredNotes
MODAL_TOKEN_IDCreate at modal.com/settings/tokens.
MODAL_TOKEN_SECRETPaired with MODAL_TOKEN_ID.
HF_TOKENOnly needed if the LeVo weights become gated; public today.

On first use the plugin downloads the weights (~13 GB checkpoint + runtime bundle) to your Modal models volume, deploys the app to your Modal account, and caches the build. Runs on an L40S GPU.