ACE-Step 1.5

The second ACE-Step generation, published in January 2026, divides songwriting between two models: a language model acts as planner, expanding a short request into lyrics, captions, tempo, key and song structure, and a diffusion transformer then renders the audio. That split covers anything from a 10-second loop to a 10-minute piece, and stays quick, finishing a full track in under ten seconds on a 3090-class card. Edits include covers guided by reference audio, regenerating a selected section, stem separation, adding layers to a mix, backing tracks for a solo vocal, and timestamped lyrics. It ships a Gradio interface, a REST server, a CLI, a Python API, portable Windows and macOS bundles, and a separate VST3 plugin for DAWs. Installation uses uv with Python 3.11 or 3.12; CUDA is the main target, though ROCm, Intel XPU, Apple MPS and CPU also work, and the 2B model runs without the planner in about 4 GB of VRAM while the 4B XL variants want 12 to 20 GB. Style LoRAs can be trained from a few songs in the UI. Code and weights are MIT, which suits producers and app builders who need a commercially usable local song generator.