ACE-Step · 2026-04-24 · notable
ACE-Step 1.5 XL — Open-Source 4B-Parameter Music Model Beats Commercial Alternatives on Consumer Hardware
Open-source 4B-parameter music model: full songs in under 2 seconds on an A100, under 4GB VRAM on consumer hardware. MIT license, 50+ languages, 1000+ instrument styles, covers/repainting/stems. Quality between Suno v4.5 and v5. 10k+ GitHub stars.
A community-built open-source music AI that outperforms most paid subscriptions — and runs locally with 4GB of VRAM.
What is it?
ACE-Step 1.5 XL is a 4B-parameter Diffusion Transformer model for music generation, released April 2, 2026, under the MIT license. It supports text-to-music, cover generation, audio repainting, stem extraction, and composition completion.
How does it work?
The model uses a hybrid architecture where a language model acts as a 'planner', transforming user queries into song blueprints (including metadata and lyrics), which are then fed to a diffusion synthesis stage. RL alignment uses intrinsic reinforcement without external reward model biases. The 'turbo' variant is distilled to generate in 8 inference steps.
Why does it matter?
ACE-Step 1.5 XL runs in under 4GB VRAM on Mac, AMD, Intel, and CUDA devices. Quality is positioned between Suno v4.5 and v5 in community testing. With Sora shutting down and Suno operating as a subscription, this represents a fully local, commercially-usable alternative. AMD officially benchmarked it on Ryzen AI hardware.
Try it
https://github.com/ace-step/ACE-Step-1.5