ACE-Step is a fast, coherent open-source foundation model for music generation. It uses a diffusion-based generation process conditioned by a lightweight linear transformer and leverages Sana’s Deep Compression AutoEncoder (DCAE) to encode audio into compact latents. It also employs representation alignment components (MERT and m-hubert), referred to as REPA, to align semantic and lyric signals during training. The main public model ACE-Step v1-3.5B delivers high throughput, synthesizing roughly 4 minutes of music in about 20 seconds on an NVIDIA A100, roughly 15× faster than several LLM-based baselines it compares to. The project supports full-song generation from natural-language prompts with duration control, lyric alignment and lyric-conditioned vocal generation, voice cloning, lyric editing, remixing, and track generation (e.g., lyric2vocal, singing2accompaniment). It includes training-free editing operations (retake, repaint, edit, extend) and fine-tuning/LoRA possibilities discussed in the community. Apache 2.0 license; no pricing; code, weights, and demos are distributed via GitHub/Hugging Face with community resources and demos.