MiniMax Releases Music 3, an Open-Weight Model for Generating Full 5-Minute Songs

What happened

MiniMax AI published MiniMax Music 3 on Hugging Face, an open-weight text-to-music model that generates complete songs — vocals, instrumentation, and arrangement — up to five minutes long from lyrics and a text description. The model weights, inference recipes, and demo are publicly available under the MiniMaxAI organization on Hugging Face.

Architecture and technical specs

Music 3 uses a hierarchical, three-part pipeline rather than a single monolithic model:

• Global LLM (8B parameters): Initialized from Qwen3-8B, this model predicts the first Residual Vector Quantization (RVQ) codebook frame-by-frame and governs long-range structure — verse/chorus progression, vocal identity, and thematic coherence across the full song.
• Local LLM (0.6B parameters): Predicts the remaining seven acoustic RVQ codebooks per frame, restoring fine-grained acoustic detail.
• Flow Matching module (2.4B parameters) and Flow-VAE Decoder (123M parameters): Fuses the hidden states from both LLMs into continuous representations, then decodes them into audio — bypassing discrete token decoding at inference time. This synthesis stack is adapted from MiniMax's speech model (MiniMax Speech), retrained for music's dynamic range.

The training tokenizer uses eight RVQ codebook layers: one semantic codebook with 16,384 entries capturing musical structure, and seven acoustic codebooks with 1,024 entries each for residual detail. Output is 32 kHz, 16-bit stereo WAV audio.

Inputs are lyrics (with optional section tags like [Verse], [Chorus], [Bridge]) and a music description covering genre, BPM, key, vocal style, and instrumentation — MiniMax recommends a structured three-part caption format for finer control.

Limitations, according to MiniMax

The model card lists several constraints: inference requires CUDA (no CPU-only path), only non-streaming generation is currently supported, text prompts are capped at 5,000 tokens, and audio generation is limited to 9,000 acoustic frames (at 25 frames per second, roughly six minutes of raw frame budget). Section tags and descriptions provide generative guidance rather than strict enforcement — the company notes outputs may not perfectly match every specified instruction.

Availability

Music 3 is supported through SGLang-Omni for serving, Hugging Face's diffusers library as a modular pipeline, and ComfyUI. MiniMax says the full-precision model fits under 24GB of VRAM, and with automatic CPU offloading plus layer-by-layer streaming of the language model, it can run on 8GB GPUs. No commercial API pricing has been published alongside the open weights; pricing not yet disclosed for any hosted version.

What this means

MiniMax is extending its speech-model architecture (Flow-VAE from MiniMax Speech) into music generation, and doing so with an open-weight release rather than a closed API — a departure from how competitors like Suno and ElevenLabs have approached music generation, largely through hosted products. The five-minute coherence target and explicit structural tagging (intro/verse/chorus/outro) address a known weak point in AI music tools: maintaining thematic and vocal consistency over long generations rather than just producing short, disjointed loops. Because weights, inference code, and a diffusers integration are all public, independent verification of MiniMax's structural-coherence and audio-quality claims should follow quickly from the open-source community — a check that closed, API-only music models rarely receive.