ACE-Step Pumps It Up With Ace-Step 1.5

    
        By vramkickedin    
     | 
    
            March 15, 2026 at 8:20 pm        
    
     | 
    
        2 min read

ACE-Step 1.5 is a new open-source music generation model that brings commercial-grade audio creation to consumer hardware. It generates full songs in under 10 seconds on an RTX 3090 while achieving quality that rivals or exceeds most commercial alternatives.

Developed jointly by ACE Studio and StepFun, this release addresses the long-standing gap between proprietary music systems and open-source tools. The previous version served primarily as a proof-of-concept, but this update delivers 'production-ready' performance that runs locally with modest hardware requirements.

Model Size: 4.79GB & VRAM GPU: <4GB required

What this model can do

Generates full songs from text prompts in over 50 languages.
Creates compositions ranging from short loops to 10-minute tracks.
Supports cover generation, repainting, and vocal-to-BGM conversion.
Enables track extraction and audio editing capabilities.
Allows users to train LoRA adapters from just a few songs for custom styles.
Produces commercially usable output with MIT license.

Musicians, producers, and content creators can use this tool to generate backing tracks, experiment with different musical styles, or create complete songs without needing expensive cloud services.

How the Ace-Step architecture works

The model uses a hybrid design that separates planning from audio generation. A Language Model acts as a composer agent, taking simple user prompts and expanding them into detailed song blueprints. According to the papers from the developers, this approach:

'fundamentally decouples usability from acoustic rendering'

and allows the audio engine to perform at a higher level.

A new distillation algorithm compresses the generation process from 50 steps down to just 4-8 steps, resulting in over 100 times faster inference speed. The team also implemented intrinsic reinforcement learning that relies on the model's internal mechanisms rather than external reward models, which eliminates certain biases common in other training approaches.