Models and their jobs - RAWCHUNKS

Rawchunks assigns each model a specific job. Audio generation runs on L4 workers. Songwriter and Composer use hosted language models; speech and image models prepare their inputs or related assets.

Music and separation

Model	Use
`acestep-v15-xl-turbo`	FAST generation, eight diffusion steps and guidance 1. Also serves the edits that support Turbo.
`acestep-v15-xl-sft`	MASTER generation, 50 diffusion steps and guidance 7. Also serves the edits that support SFT.
`acestep-v15-xl-base`	Layering, accompaniment, instrument extraction and Base editing.
`acestep-5Hz-lm-0.6B`	ACE-Step’s music language model, served through vLLM. Prepares music codes and descriptions used by the audio pipeline. The training configuration also selects this model.
`Qwen3-Embedding-0.6B`	Text conditioning for ACE-Step.
Demucs `htdemucs`	Separates vocals, drums, bass and other audio. Uses zero shifts and 0.1 overlap.

Read music options for the modes supported by each edit. A Layer or Accompany request uses Base; Rework, Extend and Restyle expose their supported alternatives.

Language and vision

Model	Use
`nvidia/Nemotron-3-Ultra-550b-a55b`	Songwriter’s planning and lyric streams; Composer’s graph-guided tool calls and pattern repair. Served by Nebius Token Factory.
`nvidia/Nemotron-3.5-Super-VL`	Converts an image’s mood, theme and visual context into a musical brief for Songwriter. Served by Nebius Token Factory.

Songwriter and Composer explain the context and outputs for each agent.

Speech and artwork

Model	Use
Cartesia `ink-whisper`	Primary transcription for spoken or sung Vibe recordings.
ElevenLabs `scribe_v2`	Fallback Vibe transcription and Editor vocal transcription. ElevenLabs forced alignment places supplied lyrics against the recording.
Google `gemini-nano-banana-2.1`	Generates separate song-cover artwork from the brief, title and style. The cover prompt uses Rawchunks’ black, acid-green, magenta and cyan palette.

Vibe audio also goes through local signal analysis, so a hum can supply musical information without transcribed words. Saved voices use ACE-Step reference conditioning and LoRA preparation.