MiniMax Music 3.0 Releases Open Weights for Five-Minute Songs

MiniMax Music 3.0 is an open-weights long-form music model

MiniMax introduced Music 3.0 on August 13, 2026 as its next-generation open-weights music-generation model. Given a creative concept and optional lyrics, the system is designed to compose, arrange, perform and render a complete song in one generation, with outputs lasting up to five minutes.

The release targets a difficult part of music generation: preserving creative intent across a full song rather than producing a short clip that drifts in instrumentation, vocal style or emotional direction. MiniMax uses fine-grained structured descriptions to represent changes in emotion, instrumentation, rhythm, vocal delivery and production style over time.

MiniMax describes the model as open weights. The announcement page does not, however, provide a simple universal license statement for every downstream use, so developers should verify the exact model repository and license terms before commercial deployment or redistribution.

A global-local language model separates structure from acoustic detail

Music 3.0 uses an 8-billion-parameter Global LLM, initialized from Qwen3.5-8B, to predict semantic and structural tokens while maintaining full-song context. A separate 0.6B Local LLM predicts acoustic tokens within each frame.

The idea is to split responsibilities: the global model tracks song-level form and long-range consistency, while the local model resolves short-range sonic detail. MiniMax says this collaboration helps maintain temporal stability over songs as long as five minutes without sacrificing variation inside individual sections.

The tokenizer uses eight-layer residual vector quantization (RVQ). The first layer captures higher-level musical semantics and structure, while seven additional layers encode progressively finer acoustic residuals.

Flow matching connects language-model features to audio rendering

Instead of sending only discrete acoustic tokens into the final decoder, Music 3.0 fuses continuous hidden states from the global and local language models. Those features condition a 2.4B flow-matching module, followed by a 123M Flow-VAE that reconstructs the final audio.

MiniMax says this design improves long-range consistency, pronunciation, instrumental coherence and fine-detail fidelity. Those quality claims come from the model developer and should be evaluated independently across genres, languages, voices, mixing styles and difficult prompts.

The announcement also describes a prompt-enhancement system that expands simple user requests into more detailed musical descriptions using terminology for genre, tempo, key, instrumentation, emotional progression and section-level changes.

The release broadens the open-weight generative-audio landscape

Open-weight music systems give researchers and developers more control over local deployment, fine-tuning and experimentation than closed music APIs. They also raise practical questions around compute requirements, training-data provenance, copyright, voice similarity, safety controls and the license terms attached to specific weights.

Music 3.0 is notable because it combines long-form generation with a hierarchical language-model architecture and continuous flow-based rendering rather than treating a song as a single undifferentiated token stream.

The five-minute duration and architecture details are part of MiniMax's official release description. They do not guarantee that every generated song will maintain quality or structural coherence for the full duration, so real-world evaluation remains necessary.