Open-source AI music generation has accelerated at a pace that would have seemed impossible two years ago. In August 2026, Chinese AI lab MiniMax released MiniMax Music 3, an open-weights model that generates complete songs up to five minutes long, with vocals, structured arrangements, and 32 kHz stereo output. It arrived just weeks after ACE-Step 1.5 claimed to beat Suno v5 on standard benchmarks and months after YuE proved that open-weight models could match commercial systems in human listening tests.
The question for anyone with a decent GPU is straightforward: which of these models deserves your VRAM? This review breaks down how MiniMax Music 3 works, how it compares to ACE-Step 1.5, YuE, Meta’s MusicGen, and Stable Audio, and where each model wins or falls short.
What Is MiniMax Music 3?
MiniMax Music 3 is an open-weights text-to-music model released on August 13, 2026, by MiniMax, a Shanghai-based AI company. The model takes two inputs: lyrics with optional section tags (like [Verse], [Chorus], [Bridge]) and a detailed music description specifying genre, BPM, key, instrumentation, and vocal style. It outputs a complete song as a 32 kHz, 16-bit stereo WAV file.
Within its first week on HuggingFace, the model accumulated over 14,000 downloads and more than 1,000 likes. The community response reflects genuine excitement: this is one of the few open-weights models that generates full songs with expressive vocals, not just instrumental loops or short clips.
The model uses a MiniMax Community License, which permits free commercial use for products generating under $20 million in annual revenue. Above that threshold, companies need written authorization from MiniMax. The license also requires displaying “MiniMax-Music3” on commercial product interfaces and includes an acceptable use policy prohibiting military applications, harmful content, and impersonation.
Under the Hood: The Hybrid-LM Architecture
MiniMax Music 3 uses a hierarchical autoregressive architecture that separates long-range musical structure from fine-grained acoustic detail. This is a fundamentally different approach from diffusion-based models like ACE-Step or pure autoregressive systems like YuE.
The Two-Language-Model Design
The architecture pairs two language models working at different scales:
- Global LLM (8B parameters): Initialized from Qwen3-8B, this model predicts the first Residual Vector Quantization (RVQ) codebook frame by frame. It handles long-range semantic structure: song sections, thematic development, vocal identity, and arrangement progression across the full duration.
- Local LLM (0.6B parameters): Predicts the remaining seven acoustic codebooks within each frame, filling in fine-grained details like timbre, texture, and harmonic nuance.
Continuous Hidden-State Synthesis
Rather than decoding only from discrete tokens, MiniMax Music 3 fuses the continuous hidden states from both LLMs. These richer representations flow through a Flow Matching module (2.4B parameters) and a Flow-VAE decoder (123M parameters), adapted from MiniMax’s speech synthesis work. The final output is 32 kHz stereo audio.
The training tokenizer uses eight layers of RVQ: a 16,384-entry semantic codebook for core musical structure, plus seven 1,024-entry acoustic codebooks for residual detail. At inference time, the system bypasses the discrete tokenizer decoder entirely, relying on the fused LLM hidden states instead.
MiniMax Music 3 vs. ACE-Step 1.5: The Open-Source Showdown
ACE-Step 1.5, released in January 2026 by ACE Studio and StepFun, is widely considered the most practical open-source music model available for creating music without licensing fees. It has published benchmark scores, runs on virtually any hardware, and carries an MIT license. Here is how the two models compare head-to-head.
| Feature | MiniMax Music 3 | ACE-Step 1.5 |
|---|---|---|
| Release Date | August 13, 2026 | January 28, 2026 |
| Architecture | Hybrid-LM (8B + 0.6B) + Flow Matching (2.4B) | LM Planner + DiT Diffusion Renderer |
| Total Parameters | ~11B | 2B to 4B DiT + 0.6B to 1.7B LM |
| Min VRAM | 8 GB (with CPU offloading) | 4 GB (INT8 quantized) |
| Full Quality VRAM | 24 GB | 8 to 16 GB |
| Max Song Duration | 5 minutes | 4 to 10 minutes |
| Vocals | Yes | Yes |
| Output Format | 32 kHz, 16-bit stereo WAV | 48 kHz stereo |
| Generation Speed | Not published | Under 2 seconds on A100 |
| Platform Support | CUDA (official); ROCm/Apple via community ports (ComfyUI, MLX, CMF, audio.cpp) | CUDA, ROCm, Intel XPU, Apple CPU |
| License | Community License ($20M revenue threshold) | MIT (fully permissive) |
| Published Benchmarks | None | SongEval 8.09, AudioBox CU 8.09 |
| LoRA Training | No | Yes (8 songs, 1 hour on RTX 3090) |
| Languages Supported | Primarily English | 50+ |
Where ACE-Step 1.5 Wins
ACE-Step 1.5 has clear advantages in accessibility and flexibility. Its 4 GB minimum VRAM means it runs on budget GPUs that cost a fraction of what MiniMax Music 3 requires. The MIT license imposes no revenue thresholds or attribution requirements. Multi-platform support extends to AMD and Intel GPUs, plus Apple Silicon via CPU inference.
The published benchmarks are significant. ACE-Step’s SongEval score of 8.09 reportedly exceeds Suno v5, and the model generates a full song in under two seconds on an A100. MiniMax Music 3 has no comparable numbers, making objective quality comparison impossible.
ACE-Step also offers six editing modes: text-to-music, cover generation, audio repainting, stem extraction, layering, and completion. MiniMax Music 3 supports generation only, with no editing capabilities.
Where MiniMax Music 3 Wins
MiniMax Music 3’s strongest argument is its architecture. The 8B Global LLM handles long-range musical structure in a way that smaller models struggle to match. User reports consistently praise the model’s ability to maintain musical coherence across five-minute songs, with recognizable verse-chorus-bridge structures and stable vocal identity throughout.
The two-input system (lyrics plus structured music description) gives composers precise control over genre, BPM, key, instrumentation, vocal style, and section-level arrangement changes. A built-in caption-rewriting tool expands short descriptions into the detailed format the model prefers.
Audio quality at 32 kHz stereo is clean and free of the artifacts that plague many open-source models, though ACE-Step’s 48 kHz output technically offers higher sampling resolution.
How Does It Compare to YuE, MusicGen, and Stable Audio?
The open-source music landscape extends well beyond these two models. Here is how MiniMax Music 3 stacks up against the other major players.
| Model | Vocals | Max Duration | Min VRAM | License | Key Strength |
|---|---|---|---|---|---|
| MiniMax Music 3 | Yes | 5 min | 8 GB | Community | Long-range song structure |
| ACE-Step 1.5 | Yes | 4 to 10 min | 4 GB | MIT | Speed, editing tools, accessibility |
| YuE | Yes | ~5 min | 16 GB | Apache 2.0 | Best KL divergence scores |
| MusicGen Large | No | 30 sec (extendable) | 8 GB | CC-BY-NC | Best published FAD/CLAP scores |
| Stable Audio Open | No | ~47 sec | 12 GB | Community | Tempo and key adherence |
YuE: The Research Favorite
YuE is a 7B autoregressive model from M-A-P that achieved the best KL divergence score (0.372) among tested systems, beating both Udio (0.503) and Suno V4 (0.620). In human evaluation with 40 raters, YuE tied with Tiangong and Udio on preference and musicality, trailing only Suno V4.
The downside is speed: YuE runs approximately 12 times slower than realtime on an RTX 4090. A five-minute song takes roughly an hour to generate. MiniMax Music 3, while slower than ACE-Step, does not have this limitation. YuE also requires 16 to 24 GB of VRAM, putting it in the same hardware tier as MiniMax Music 3.
Meta MusicGen: The Instrumental Specialist
MusicGen Large remains the strongest published performer on standard benchmarks, with an FAD of 2.82 and CLAP score of 0.51. However, it generates instrumental music only, with a native duration of 30 seconds that requires extension techniques for longer compositions. Its CC-BY-NC license prohibits commercial use of the weights. For anyone who needs vocals or full-song structure, MusicGen is not a direct competitor.
Stable Audio: The Tempo Controller
Stable Audio 3 excels at adhering to specified tempo (61% accuracy) and key (64% accuracy), making it the most reliable choice for producers who need backing tracks at exact specifications. However, it generates short clips rather than full songs and does not produce vocals.
Running MiniMax Music 3 at Home
The practical question for most readers: can your hardware handle it?
VRAM Requirements
- 24 GB or more: Full precision generation with no offloading. This means an RTX 4090, RTX A5000, or better.
- 22 GB: Automatic CPU offloading reduces peak usage to approximately 22 GB with bfloat16 precision. An RTX 3090 or RTX 4080 Ti can handle this.
- 8 GB: With group offloading that streams the language model layer by layer, the model fits on 8 GB cards like the RTX 4060 or RTX 3070. Generation is significantly slower.
Community-provided 8-bit and 4-bit quantized versions further reduce the footprint, allowing the model to run on substantially lower VRAM than the full-precision build at the cost of some quality trade-offs. These quantized builds are especially useful when combined with offloading on consumer cards and with the unofficial Apple and AMD ports described below.
The CUDA Problem and Community Workarounds
Officially, MiniMax Music 3 requires NVIDIA CUDA hardware, with no native support for AMD GPUs, Intel GPUs, or Apple Silicon. Compared to ACE-Step 1.5, which runs natively on CUDA, ROCm, Intel XPU, and Apple CPU, that is a significant limitation on paper.
In practice, the community has already bridged much of that gap. Users report successfully running MiniMax Music 3 on AMD hardware via ROCm through ComfyUI, and a dedicated CMF version for ComfyUI is reported to work on Apple M4 and M5 systems, with potential compatibility for AMD and Intel hardware as well. An MLX port also enables inference on Apple Silicon outside the official CUDA stack. Separately, discussion on Hugging Face notes that audio.cpp now supports the model, which may open further cross-platform options. Vulkan support, however, still appears to be lacking, so coverage is not yet universal.
Setup Options
The model supports three inference frameworks:
- SGLang-Omni: The recommended server approach, using a shared speech API compatible with OpenAI-style endpoints.
- Diffusers: HuggingFace’s modular pipeline, suitable for experimentation and integration with existing workflows.
- ComfyUI: A node-based interface with community-built workflows for visual music generation.
Installation requires downloading approximately 10 GB of model weights from HuggingFace. A prompt-enhancement tool can expand short descriptions into the structured format the model performs best with.
What MiniMax Music 3 Gets Right
The model’s standout qualities are structural coherence and vocal expressiveness. Several aspects deserve specific mention.
- Full-song structure: The model reliably produces songs with recognizable intros, verses, pre-choruses, choruses, bridges, instrumental breaks, and outros. This is something most open-source models struggle with.
- Vocal identity stability: The singer’s voice remains consistent across sections, avoiding the timbral drift that affects models like YuE.
- Detailed control: The structured caption system lets users specify BPM, key, instrumentation changes per section, vocal characteristics, and production style.
- Audio quality: The 32 kHz stereo output is clean, with good separation between vocals and instruments. The Flow-VAE synthesis path, adapted from MiniMax’s speech work, produces natural vocal articulation.
- Active community: Over 18 HuggingFace Spaces, 11 finetunes, 20 quantizations, and 3 adapters appeared within the first week, suggesting rapid ecosystem development.
Where MiniMax Music 3 Falls Short
No model is perfect, and MiniMax Music 3 has meaningful limitations.
- No published benchmarks: MiniMax has not released FAD, CLAP, MOS, or SongEval scores. All quality claims come from company-selected demo samples. Without independent benchmark data, comparing its output quality to ACE-Step or YuE on objective metrics is impossible.
- Limited official platform support: Officially CUDA-only, which locks out AMD, Intel, and Apple hardware without community ports. Unofficial options now exist via ComfyUI with ROCm for AMD, an MLX build for Apple Silicon, a CMF build reported to run on Apple M4/M5, and audio.cpp support, but Vulkan support is still lacking and these ports are not officially supported by MiniMax.
- Higher VRAM floor: While 8 GB is technically possible with aggressive offloading, the practical minimum for reasonable generation speed is closer to 22 GB. ACE-Step runs well on 4 to 8 GB.
- No editing capabilities: The model generates complete songs but offers no repainting, stem extraction, or completion features. ACE-Step provides all six editing modes.
- Non-English weakness: Independent testing found the model struggles significantly with non-Western genres. Bollywood and Colombian cumbia prompts produced results that failed to capture genre-specific rhythmic and melodic conventions.
- Non-streaming generation: The model cannot produce audio incrementally. Users must wait for the full song to complete before hearing any output.
- Restrictive license: The Community License’s $20 million revenue threshold and attribution requirements are more limiting than ACE-Step’s MIT license or YuE’s Apache 2.0. Companies approaching that revenue level need to plan for licensing negotiations.
Frequently Asked Questions
Is MiniMax Music 3 free to use?
Yes, for both personal and commercial use, provided your product generates less than $20 million in annual revenue. Above that threshold, you need written authorization from MiniMax. The license also requires displaying “MiniMax-Music3” on your product’s user interface and prohibits military use, harmful content, and several other categories listed in the acceptable use policy.
Is there a better music AI than Suno?
Among open-source models, none currently match Suno v5’s overall quality in independent listening tests. The closest contenders are ACE-Step 1.5 (which claims competitive benchmark scores) and YuE (which tied with Udio in human evaluation). MiniMax Music 3 has not published benchmark data, so its position relative to Suno remains unclear. For instrumentals, Stable Audio 3 outperforms all other open models on tempo and key adherence.
How much does MiniMax Music 3 cost to run?
The model is free to download and run locally. Hardware costs vary: an RTX 4090 (24 GB VRAM) costs approximately $1,500 to $2,000, while an RTX 3090 (24 GB) can be found used for $600 to $800. For the 8 GB offloading path, an RTX 4060 costs around $300. Cloud GPU rental through services like RunPod or Vast.ai typically costs $0.30 to $0.60 per hour for a 24 GB card.
Is AI-generated music legal?
AI-generated music is legal to create and distribute in most jurisdictions, but the legal landscape is evolving. MiniMax Music 3’s Community License explicitly grants commercial use rights for generated outputs. However, copyright protection for AI-generated works varies by country. In the United States, the Copyright Office has indicated that purely AI-generated content may not be eligible for copyright protection, though human creative contributions to the prompt or arrangement could change that analysis.
Can MiniMax Music 3 run on a Mac?
Officially, no. MiniMax only supports NVIDIA CUDA, so there is no official Apple Silicon build. Unofficially, the community has released an MLX version for Apple hardware and a CMF version for ComfyUI reported to run on Apple M4 and M5 chips, and Hugging Face discussion notes that audio.cpp now supports the model as well. These options are community-maintained and may also help on AMD and Intel hardware, though Vulkan support still appears to be lacking. For guaranteed native support, ACE-Step 1.5 remains the safer choice with official Apple CPU inference.
What is the difference between MiniMax Music 3 and MiniMax Music 2.5?
MiniMax Music 3 is a significant architectural upgrade. It uses a larger 8B Global LLM (initialized from Qwen3-8B) compared to Music 2.5’s smaller models, introduces the Flow Matching synthesis path for higher audio quality, and supports five-minute song generation with finer-grained structural control through section tags and structured captions.
The Bottom Line
MiniMax Music 3 is a genuinely capable music generation model that excels at producing structurally coherent, vocally expressive songs. Its hierarchical architecture, separating long-range composition from frame-level acoustics, is an innovative approach that yields impressive results on English-language pop and similar genres.
However, it enters a competitive field where ACE-Step 1.5 sets a high bar for accessibility, speed, and licensing freedom. ACE-Step runs on hardware that costs a fraction of what MiniMax Music 3 requires, supports every major GPU platform, carries an MIT license, and has published benchmark scores that reportedly beat commercial systems.
MiniMax Music 3’s biggest gap is the absence of independent quality benchmarks. Without published FAD, CLAP, or SongEval scores, users must rely on company demos and community reports to assess quality. This is a notable omission for a model competing in a space where ACE-Step and YuE have transparent evaluation data.
If you have 24 GB of NVIDIA VRAM and want the best long-form song structure with vocals, MiniMax Music 3 is worth testing. If you need lower hardware requirements, editing capabilities, multi-platform support, or a permissive license, ACE-Step 1.5 remains the safer and more flexible choice. For research purposes where audio quality metrics matter, YuE’s published benchmark data makes it the more verifiable option.
The open-source music AI landscape is moving fast. MiniMax Music 3 is a strong addition to the field, but it is not yet the clear winner its demo page might suggest. The best model for your use case depends on your hardware, your license requirements, and whether you need benchmarks you can trust.
