MiniMax releases MiniMax-Music3 open-weights music AI

Chinese AI developer MiniMax has released MiniMax-Music3, an open-weights music generation model capable of creating complete tracks up to five minutes long from text descriptions. A key technical advancement of this model is its ability to maintain structural coherence and vocal consistency over longer durations, a common issue in previous text-to-music models.

The architecture utilizes two distinct language models: an 8-billion parameter global model to manage semantic progression and a 0.6-billion parameter local LLM to handle acoustic details frame-by-frame. These models work with a 2.4-billion parameter Flow Matching module to produce high-quality WAV stereo audio.

The tool supports 40 languages, including Italian and Japanese. Because it is an open-weights model, it can be run locally—such as through ComfyUI—allowing for unlimited generation without reliance on cloud services. The release also includes tools to help users transform short inputs into structured prompts and long lyrics optimized for the model.