Editorial illustration for MiniMax-Music3 Generates Five-Minute Songs from Lyrics and Captions
MiniMax-Music3 Generates 5-Min Songs from Lyrics
MiniMax shipped MiniMax-Music3 on the same day it announced the model, with weights, inference code and three serving paths ready to use immediately. Feed it lyrics with section tags and a separate structured caption describing the sound, and it returns a full song up to five minutes long in one pass, rendered as 32 kHz, 16-bit stereo WAV. That's a single generation covering intro, verses, chorus and outro, not stitched loops.
The release matters for who can actually use it Monday morning. Solo musicians, indie game studios and small marketing teams get a model they can deploy without waiting on API access or a research paper turning into a product. MiniMax's Community License allows commercial use but comes with conditions: attribution in the product UI, a $20 million annual revenue threshold that triggers a separate licensing conversation with MiniMax, and a requirement that anyone hosting third-party generations build in safeguards against infringing output.
Under the hood, the system splits work between a Hybrid-LM handling structure and a continuous synthesis stack handling audio quality. How those pieces divide the labor, and what that means for fidelity, is where the architecture gets interesting.
MiniMax released MiniMax-Music3, an open-weights text-to-music model. The model takes two separate inputs: lyrics carrying section tags, and a detailed music description. It returns a complete song of up to five minutes in a single generation, as 32 kHz, 16-bit stereo WAV.
Why this matters
MiniMax just handed developers a full production stack, not a demo. Open weights plus inference code plus three serving paths means teams can self-host, fine-tune, or hit an API depending on budget, and that flexibility matters more than the five-minute runtime itself. The listed use cases, UGC background scores, game music, ad beds, are exactly where per-song generation cost has kept AI music out of production pipelines. If MiniMax-Music3 holds up outside curated demos, it undercuts the closed-model incumbents on price and control at once.
We'd still want independent listening tests before betting a product on it. Structure and vocal coherence over five minutes is where most music models fall apart, looping sections or losing the melodic thread past the two-minute mark. The Hybrid-LM approach (8B global, 0.6B local) plus Flow-VAE is a real architectural bet worth watching, but architecture claims and perceived audio quality are different things.
For founders building on this, the open weights de-risk vendor lock-in. For researchers, the interesting question is whether the section-tag lyrics input generalizes past English or scales to genres outside whatever MiniMax trained on.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives OpenAI Dissolves AI Risk Team, Splits Duties to Existing Divisions
From Archives OpenAI Shuts Down Team Tasked With Evaluating AI Catastrophic Risks
Common Questions Answered
What are the two separate inputs required for MiniMax-Music3 to generate a song?
MiniMax-Music3 requires lyrics with section tags and a detailed structured caption describing the desired sound. These two inputs work together to guide the model in generating a complete, coherent song that matches both the lyrical content and musical style specifications.
What is the maximum length and audio format of songs generated by MiniMax-Music3?
MiniMax-Music3 can generate complete songs up to five minutes long in a single pass, rendered as 32 kHz, 16-bit stereo WAV files. This full-length generation covers the entire song structure including intro, verses, chorus, and outro without requiring stitched loops.
How does MiniMax-Music3's open-weights release benefit developers compared to closed-source alternatives?
By providing open weights, inference code, and three serving paths, MiniMax-Music3 gives developers flexibility to self-host, fine-tune, or use an API depending on their budget and needs. This production-ready stack enables teams to integrate AI music generation into real workflows for use cases like UGC background scores, game music, and ad beds where per-song generation costs previously prevented adoption.
What makes MiniMax-Music3 different from previous music generation approaches that used stitched loops?
MiniMax-Music3 generates an entire five-minute song in a single pass rather than combining multiple stitched loops together. This unified generation approach creates more coherent and professionally structured songs with proper transitions between sections like intros, verses, choruses, and outros.
What production use cases is MiniMax-Music3 designed to address?
MiniMax-Music3 targets production pipelines for user-generated content background scores, game music, and advertising beds. The model's ability to generate full songs in one pass at lower costs aims to make AI music generation economically viable for these previously cost-prohibitive applications.
Further Reading
- MiniMax Music 3.0: Next-Generation Open-Weights, Production-Ready Versatile Music Model - MiniMax
- MiniMax releases Music 3 for five-minute songs with open weights and structured lyrics input - RuntimeWire
- MiniMax releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption - MarkTechPost
- MiniMax Music 3 in ComfyUI: AI Music Generation - ComfyUI Docs
- MiniMax has released 'MiniMax-Music3,' a music generation AI, for free, following its video generation AI, allowing users to generate songs with Japanese vocals up to 5 minutes in length locally. - GIGAZINE
