MiniMax drops open-weights model crafting 5-minute songs

MiniMax Music 3.0 is a new open-weights music generation model that crafts fully arranged, ready-to-listen songs up to five minutes long from a simple text prompt and optional lyrics.

Most AI audio generators struggle with length. Give them a prompt, and specific instruments vanish after two verses, or vocals lose their style halfway through. Music 3.0 tackles that structural drift head-on, producing clean instrumentation and natural vocal delivery in a single pass.

What it does

Feed it a genre, tempo, musical key, and lyrics tagged with parts like verses, choruses, and bridges, and Music 3.0 produces a completed production. It handles genres ranging from vintage jazz and stadium pop rock to 126 BPM progressive house.

To bridge the gap between casual ideas and studio vocabulary, it includes a Prompt Enhancement System. This automatically translates plain creative briefs into Structured Captions, instructing the model on instrument entrances, low-end grooves, and vocal intensity curves across the song.

Best of all, you do not need an enterprise server to run it. Community testing shows the entire model weights measure around 9.8 GB, easily fitting into mid-tier GPUs. An int8 quantized build drops hardware requirements even further, running smoothly on entry-level setups.

How it works

The architecture relies on a three-tier pipeline: a hierarchical tokenizer, a dual-layer language model, and a continuous flow-matching synthesis engine.

Instead of dumping everything into one token layer, an eight-layer residual vector quantization scheme isolates core song semantics on the first layer, using the remaining seven layers to encode acoustic residuals.

At the center sits a Hybrid-LM. An 8B Global LLM, initialized from Qwen3.5-8B, predicts frame-by-frame structural tokens across the entire timeline. Working alongside it, a 0.6B Local LLM resolves fine-grained acoustic tokens within each frame.

For final rendering, Music 3.0 skips standard discrete token decoding. It fuses continuous hidden states from both language models into a 2.4B flow-matching module, feeding a 123M Flow-VAE to reconstruct high-fidelity audio without the usual digital artifacts.

Why it matters

For indie creators, game developers, and tinkerers, Music 3.0 is a massive win for local production. Having an open-weights generator capable of handling five continuous minutes means no more awkward stitching or subscription lock-in.

We think MiniMax made a clever trade-off here. By pairing a Qwen3.5-8B backbone with continuous flow matching, they packed production-ready sound into a sub-10 GB footprint. It punches well above its weight class, though dialing in exact instrumental nuance will still take some experimenting with its structured prompting tools.