MiniMax has released MiniMax-Music3, an open-weights text-to-music model that generates complete five-minute songs in a single pass, with full code and model weights available for deployment starting August 13, 2026. The model accepts two inputs: lyrics tagged with song sections like verse and chorus, plus a detailed music description. It outputs high-quality stereo audio at 32 kilohertz, 16-bit resolution, making it immediately usable for creators and developers rather than a research-only preview.

How Does MiniMax-Music3 Actually Generate Music?

The architecture combines two separate language models with a continuous synthesis pipeline. A larger 8-billion-parameter global language model predicts the overall song structure and long-range musical coherence, while a smaller 600-million-parameter local model handles fine-grained acoustic details within each time step. Rather than decoding from discrete tokens like many music AI systems, MiniMax fuses the outputs of both models and feeds them into a 2.4-billion-parameter flow-matching module, which maps the result into a latent space decoded by a 123-million-parameter Flow-VAE (variational autoencoder). This continuous synthesis approach, inherited from MiniMax's speech technology, skips the discrete tokenizer decoder entirely at inference time.

The training process uses eight layers of residual vector quantization. The first semantic codebook contains 16,384 entries and captures core musical structure, while seven acoustic codebooks with 1,024 entries each encode residual detail. Training optimizes the semantic layer first, then all eight layers jointly, allowing the model to learn hierarchical musical representations.

What Industries and Use Cases Benefit Most?

MiniMax-Music3 targets a broad range of applications across multiple sectors. The model is designed for game developers creating adaptive background music that changes based on gameplay, advertising agencies producing localized sonic branding, short-form video creators needing background scoring, e-learning platforms adding audio to educational content, podcasters generating intro and outro music, fitness app developers creating mood-conditioned playlists, and retail environments generating in-store audio. Songwriters can use it to generate scratch and demo tracks, while music-tech software-as-a-service (SaaS) companies can integrate it for offline batch generation where per-song API costs are a constraint.

• Game Development: Adaptive music that responds to level progression and player actions without requiring manual composition for each scenario.
• Advertising and Branding: Localized sonic branding and ad beds tailored to regional markets and brand guidelines.
• Creator Tools: Background scoring for user-generated content videos and short-form social media, reducing production friction.
• Music Production: Scratch tracks and demo versions for songwriters to iterate on before recording with live musicians.
• Batch Processing: Offline generation for applications where per-song API costs are prohibitive, such as large-scale playlist generation.

What Are the Hardware and Deployment Requirements?

MiniMax shipped three documented serving paths to accommodate different hardware constraints. The reference implementation uses SGLang-Omni, a specialized inference server that requires two CUDA graphics processing units (GPUs), with one running the language models and RVQ (residual vector quantization) autoregressive generation and the other running flow matching and audio decoding. For developers with more limited hardware, the diffusers modular pipeline fits under 24 gigabytes of VRAM (video random-access memory) at full precision, approximately 22 gigabytes with automatic CPU offload, and down to 8 gigabytes with leaf-level group offloading. A ComfyUI native template using repacked FP16 and INT8 weights from Comfy-Org provides a third option for users preferring node-based workflows.

The accessibility of these deployment options matters significantly. Solo creators and indie studios can run the model on consumer-grade hardware with careful optimization, mid-market teams can deploy via cloud infrastructure, and large organizations can integrate it into existing music-tech platforms. This democratization contrasts with earlier AI music models that required substantial computational resources or were available only through proprietary APIs.

What Are the Licensing Terms and Commercial Restrictions?

MiniMax-Music3 operates under a Community License that permits commercial use with specific conditions. Any product using the model must display "MiniMax-Music3" prominently in the user interface. Organizations whose aggregate yearly revenue from products using the model exceeds $20 million must obtain separate written authorization from MiniMax before deployment. Additionally, anyone hosting third-party music generation must implement and maintain safeguards against infringing outputs, such as music that violates copyright or mimics existing copyrighted works without permission.

This licensing structure sits between fully open-source models and proprietary commercial services. It allows small creators and startups to experiment and build products without upfront licensing fees, while ensuring MiniMax retains control over large-scale commercial deployments. The requirement to implement safeguards against infringing outputs reflects ongoing industry concerns about AI music systems potentially generating music that violates existing copyrights.

How Does MiniMax-Music3 Compare to Other AI Music Approaches?

The release of MiniMax-Music3 reflects a broader shift in AI music generation toward open-weights models and continuous synthesis methods. Earlier AI music systems often relied on discrete token prediction, similar to language models, which can introduce artifacts and limit audio quality. MiniMax's hybrid approach of combining discrete semantic tokens with continuous synthesis aims to preserve musical coherence while capturing fine acoustic detail. The immediate release of weights and inference code also contrasts with some competitors that release models through proprietary APIs or with significant delays between announcement and availability.

The model's ability to generate complete five-minute songs in a single pass, rather than requiring iterative refinement or stitching together shorter segments, represents a technical advancement. Most earlier AI music systems generated shorter clips or required post-processing to create full-length tracks. MiniMax-Music3's architecture, with its hierarchical language models and flow-matching synthesis, enables end-to-end generation of longer compositions.

What Do the Input Requirements Look Like in Practice?

Users provide two structured inputs to MiniMax-Music3. Lyrics carry the words and section tags on separate lines, including standard song structure markers like [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], and [Outro]. A separate Structured Caption provides three components: Global Metadata (tempo, key, overall mood), Vocal Details (vocal style, gender, tone), and Arrangement (instrumentation, production style). MiniMax also ships a music-caption-rewriter agent skill that expands a short text description into the full three-part format offline, reducing friction for users unfamiliar with structured prompting.

This structured input design reflects lessons from earlier AI music systems. Rather than relying on free-form natural language descriptions that can be ambiguous, MiniMax enforces a consistent format that captures the key musical dimensions. The caption-rewriter agent bridges the gap between casual user input and the structured format, allowing both technical and non-technical users to generate music effectively.

What Happens When AI Music Meets Physical Media?

While MiniMax-Music3 focuses on digital music generation, a parallel development shows how AI audio compression is enabling entirely new forms of music distribution. A maker known as Makestreme recently printed an entire two-minute song onto a sheet of paper using Meta's EnCodec neural audio codec, compressing a 2.9-megabyte MP3 file down to roughly 21 kilobytes, a reduction of approximately 1,000 times. The compressed tokens fit across eight QR codes arranged as a double-sided "paper cassette".

The EnCodec codec, released by Meta in 2022, converts audio waveforms into discrete tokens that a matching decoder reconstructs back into audio. At the project's 3 kilobits-per-second setting, the reconstructed song reportedly held up well, though at 1.5 kilobits-per-second it produced heavily degraded output. Each QR code carries a one-byte ID plus a chunk of the token stream, with four codes per side in a nod to flipping a vinyl record. The generator script uses the lowest error correction level to maximize density, meaning there is no redundancy to spare; damage one code and the song is lost.

This experiment highlights a fundamental challenge for AI-generated and AI-compressed audio: longevity depends on the neural network that decodes it. The paper may last centuries, but the song only survives as long as the EnCodec decoder remains available and functional. This creates a preservation paradox for AI music, where the physical medium outlasts the computational infrastructure required to play it.

MiniMax-Music3's release into the open-weights ecosystem addresses this concern partially by ensuring the model weights remain available for future deployment, even if the original company discontinues support. However, the broader question of how to preserve AI-generated and AI-compressed music for long-term access remains unresolved across the industry.