ACE-Step 1.5

ACE-Step 1.5 is an open music model from ACE Studio and StepFun that writes a full song from a style description and lyrics, with vocals. Per the model card, it supports 50+ languages and tracks up to 10 minutes, and offers cover, repaint, vocal-to-accompaniment and reference-audio modes, plus LoRA training. A language-model planner (0.6B, 1.7B or 4B, based on Qwen3) works together with a diffusion transformer (2B; the XL series has 4B). The authors state that a song takes under 2 seconds on an A100 and under 10 seconds on an RTX 3090. Version 1.5 came out on 2026-01-23, the XL series on 2026-04-02. The weights are public on Hugging Face without any acceptance step, under the MIT license; the authors say the music can be used commercially and name licensed, royalty-free and synthetic training data, which we have not verified. The card states under 4 GB of VRAM; the repository suggests 8–16 GB for the 2B models and at least 12 GB with offload and quantization for XL. The authors offer an official demo Space on Hugging Face that you can try without installation; code is on GitHub and the paper is on arXiv. Date of check: 2026-10-18.