A practical guide to installing MiniMax Music 3 locally, covering VRAM needs, the Gradio setup, and what the open music model actually delivers.
What is MiniMax Music 3? #
MiniMax Music 3 is an open text-to-music model from MiniMaxAI that generates complete songs, vocals included, from a lyric sheet and a text description of the sound you want. It’s built to handle full song structure (intro, verse, pre-chorus, chorus, bridge, instrumental break, outro) rather than short loops or isolated clips, and outputs tracks up to five minutes long as 32 kHz, 16-bit stereo WAV. You can install it on a local machine with a capable GPU using the Hugging Face weights and a diffusers
-based script, with a Gradio interface layered on top for prompt input and playback.
TL;DR #
MiniMax Music 3 generates full songs with vocals, structured arrangements, and stable audio quality across tracks up to five minutes long.- The model uses a hierarchical auto-regressive design: an 8B-parameter global language model handles long-range song structure, while a smaller ~0.6B local model fills in acoustic detail frame by frame. - Local installation runs through Hugging Face and diffusers, with a Gradio web UI for entering lyrics, genre, BPM, key, and vocal details. - On an RTX 6000 with 48GB of VRAM, generation consumed just over 27GB, meaning it should also run on high-end 32GB consumer cards depending on batch settings. - A Studio mode in the Gradio demo gives finer control over global metadata: vocal timbre, instrumentation, section-level arrangement, and song title. - Testing showed strong results in English but noticeably weaker output forHindi/Bollywood-style and Spanish/Latin genres like cumbia, suggesting multilingual and genre coverage outside English pop is inconsistent. - The model relies on an eight-layer RVQ tokenizer,** flow matching**, and a** flow variational autoencoder**adapted from MiniMax’s speech model to turn latents into final waveforms.
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How do you install MiniMax Music 3 locally? #
The setup follows a fairly standard Hugging Face model workflow. On a Linux system (the reference setup here used Ubuntu with an Nvidia RTX 6000, 48GB VRAM), the core steps are:
- Install
diffusers
andtorchaudio
, since MiniMax Music 3 ships as diffusers-compatible safetensors weights alongside a dedicated language model component. - Pull the model repository from Hugging Face (
MiniMaxAI/MiniMax-Music3
), which includes the language model, a Qwen3-8B tokenizer used for music conditioning, a flow-matching VAE, an RVQ depth decoder, a vocoder, and a transformer checkpoint split across safetensors shards. - Run the provided end-to-end script, which downloads the model weights on first launch and starts a Gradio demo server.
- Open the Gradio interface in a browser to enter lyrics, genre, and structural descriptions, then generate audio.
The repository ships with a working test script (scripts/end_to_end/minimax_ttm_test.py
) that developers can adapt, and a Gradio wrapper makes it usable without touching Python code directly once installed. No conda environment is strictly required, standard Python virtual environments work fine.
How much VRAM does MiniMax Music 3 need? #
In testing on an RTX 6000 with 48GB of VRAM, the model consumed just over 27GB during generation with no additional batching or optimization applied. That leaves meaningful headroom on a 48GB card and suggests the model could fit on 32GB professional or high-end consumer GPUs, though there’s no confirmed lower bound from official documentation. The model’s file structure gives a sense of why it’s VRAM-hungry: it bundles an 8B-parameter language model, a separate ~0.6B local model, a Qwen3-8B tokenizer/conditioning stack, a flow-matching VAE, a vocoder, and a diffusion transformer, all of which need to be resident in memory simultaneously during inference.
For anyone without a card in that range locally, renting GPU time from a cloud provider is a common workaround, since the model doesn’t require training, only inference-time VRAM.
How does the Gradio interface work? #
The demo exposes two main modes. The simple mode takes a text prompt describing genre, tempo, key, mood, and instrumentation (for example: “acoustic pop, C major, warm and building”), and the model writes its own lyrics and generates the full track based on that description alone.
Studio mode gives more granular control. Here you supply your own lyrics, and the interface breaks them into song sections (verses, chorus, bridge) for structured generation. It also exposes fields for vocal detail (gender, tone descriptors like “breathy tenor”), instrumentation choices, section-level arrangement notes, and a song title. This mode is closer to how a producer would brief a session: define the skeleton, then let the model fill in the performance.
Generation isn’t instantaneous. Depending on hardware and network conditions (especially on rented cloud GPUs), a full track can take a couple of minutes to render once the prompt is submitted.
What does the architecture actually do? #
MiniMax Music 3 uses a hierarchical auto-regressive design that splits the generation task into two coupled language models:
- A global LLM(8 billion parameters, initialized from a 1.3B base) predicts the first RVQ (residual vector quantization) codebook frame by frame. This model owns the song’s long-range semantic and structural progression, meaning it’s responsible for keeping the song coherent across verses, choruses, and transitions over multiple minutes. - A local LLM(around 0.6 billion parameters) fills in the remaining acoustic codebooks within each frame, restoring the fine-grained detail that gives vocals and instruments their texture.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Rather than decoding purely from discrete tokens, the synthesis pipeline fuses continuous hidden states from both models before handing off to audio synthesis, preserving more information for vocal articulation and instrumental texture. An eight-layer RVQ tokenizer underpins training, and the model uses flow matching, a generative technique that transforms noise into target audio latents by following a smooth, continuous path, rather than the discrete diffusion steps used in some competing systems. A flow variational autoencoder, adapted from MiniMax’s speech model and retrained on music, decodes the final latent into a stereo waveform.
Is MiniMax Music 3 worth running locally? #
It depends heavily on what you’re generating. English-language pop and similar genres, based on testing with prompts like “acoustic pop, C major, warm and building,” produced coherent full songs with structured verses, choruses, and reasonably stable vocal quality across the track. The model clearly delivers on its core promise there: full song generation with evolving arrangement rather than a short repeating loop.
Results were much weaker outside that lane. A Bollywood-style item song prompt, using Hindi-inspired lyrics and vocal metadata, produced output that fell far short of the genre’s actual sonic identity, lacking the vocal character and instrumentation associated with well-known Bollywood composers. A cumbia prompt (a genre with roots in Colombian folk music blending indigenous, African, and Spanish influences) similarly produced English-language vocals over a rhythm track that didn’t capture the accordion-and-brass-driven groove associated with the style.
The model card doesn’t explicitly confirm multilingual training, and based on hands-on testing, quality outside English and Western pop structures appears inconsistent. For developers building English-language music tools, that’s a reasonable tradeoff. For anyone targeting regional genres or non-English vocals, it’s worth generating test samples before committing to it as a production tool.
Frequently Asked Questions #
What GPU do you need to run MiniMax Music 3?
Testing showed roughly 27GB of VRAM used on an RTX 6000 (48GB) during generation. A 32GB or larger GPU is a reasonable starting point, though no official minimum has been published.
Can MiniMax Music 3 generate songs in languages other than English?
It attempts to, but quality drops noticeably outside English. Tests with Hindi-style Bollywood lyrics and Spanish-influenced cumbia prompts produced vocals and instrumentation that didn’t match the target genre’s actual sound.
How long are the songs it generates?
The model is built to produce structurally coherent full songs up to five minutes long, complete with intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro sections.
What audio format does it output?
MiniMax Music 3 outputs 32 kHz, 16-bit stereo WAV files.
Do you need to write full lyrics yourself?
No. In simple mode, you can give it just a genre and mood description and it will write its own lyrics and structure. Studio mode lets you supply your own lyrics and control section-by-section arrangement and vocal characteristics.