Can MiniMax Music 3 Keep Unreleased Lyrics Local? What The 57 GB Download Actually Proves

August 14, 2026

MiniMax Music 3 can run locally, but a 57 GB download and an 8 GB GPU path do not prove lyrics stay private. Here is how to test the full workflow.

Yes, MiniMax Music 3 can keep unreleased lyrics, music descriptions, and generated audio on hardware you control. MiniMax released downloadable weights and documents local inference through SGLang-Omni, Diffusers, and ComfyUI. Its example sends lyrics to a loopback address and saves a WAV file locally instead of requiring MiniMax's hosted API.

But that proves the model can run locally—not that every installation is private. The current repository totals about 57.35 GB, the runtime needs CUDA, and downloads, package installation, logs, plugins, backups, remote access, and optional tools can all create separate data paths. A browser interface saying “localhost” is useful evidence, but it is not a no-egress audit.

For creators, the practical rule is simple: use synthetic lyrics to test the entire workflow, monitor outbound connections, find every copy of the prompt and output, and only then decide whether the system is suitable for an unreleased song.

What Is Confirmed

MiniMax's official Music 3 model card describes a text-to-music model that accepts lyrics plus a detailed music description and can generate complete songs up to five minutes long. It produces 32 kHz, 16-bit stereo WAV audio.

The published architecture includes:

• An 8-billion-parameter Global LLM initialized from Qwen3-8B for long-range structure.
• A 0.6-billion-parameter Local LLM for frame-level acoustic detail.
• A 2.4-billion-parameter Flow Matching module.
• A 123-million-parameter Flow-VAE decoder.
• Eight residual-vector-quantization codebooks during training.

The model card documents three local execution paths: SGLang-Omni, Hugging Face Diffusers, and ComfyUI. Its SGLang example serves the model on 127.0.0.1, accepts lyrics and instructions through a local speech-compatible endpoint, and writes the result to a WAV file.

The current repository file tree contains 88 files totaling 57,353,379,600 bytes—about 57.35 GB in decimal units or 53.41 GiB. That total includes model components, tokenizer data, configuration, code, and an example audio file.

The hardware story changed quickly during launch. The final model card says full-precision loading fits under 24 GB of VRAM, automatic CPU offloading uses about 22 GB, and slower layer-by-layer streaming can fit on an 8 GB GPU. CUDA is still required. “Fits in 8 GB VRAM” therefore does not mean an 8 GB download, an instant setup, or fast generation.

MiniMax also published the source repository, and official installation links point to the serving frameworks rather than requiring a MiniMax cloud generation endpoint.

What Is Still Unclear

No official source reviewed for this article publishes a completed network-egress audit for every supported Music 3 workflow. MiniMax's loopback example is evidence that local inference is possible. It does not establish that every dependency, launcher, UI, plugin, update checker, crash reporter, or community wrapper remains silent on the network.

The model card does not provide comparable generation speeds for the under-24 GB, 22 GB offload, and 8 GB streaming configurations. Moving layers between system memory and the GPU can make a model fit while adding substantial latency. The minimum usable system RAM, free disk space, load time, and generation time will depend on the chosen runtime and song duration.

It is also too early to treat launch examples as independent proof of consistent output quality. MiniMax says section tags and music descriptions provide generative control rather than strict guarantees. Tempo, key, instruments, lyrics, and song structure may not match every request exactly.

The downloadable release does not answer broader rights questions for a particular song. Keeping lyrics local does not determine whether a prompt or output infringes someone else's copyright, publicity rights, contractual duties, or confidentiality obligations. The model's license expressly leaves those responsibilities with the user.

Finally, “unreleased” does not always mean “secret.” A lyric can already exist in a synced notes app, cloud backup, messaging thread, shared project, source-control repository, or collaborator's inbox before it reaches the model. The Music 3 runtime is only one part of the data path.

Does MiniMax Music 3 Upload Your Lyrics?

The downloadable model does not need to send a generation request to MiniMax's hosted service. In the official SGLang example, the request goes to http://127.0.0.1:8000/v1/audio/speech, a loopback address on the same computer. The Diffusers example instantiates the model components and generates audio in the local Python process.

That makes private inference technically possible. It does not justify a universal claim that Music 3 never uploads lyrics.

The answer for a real installation depends on what the operator runs around the model:

• Hugging Face downloads contact Hugging Face during acquisition.
• Python, Git, package managers, containers, and model managers may contact their own registries.
• A third-party web UI may add analytics, hosted prompt enhancement, remote queues, or cloud storage.
• ComfyUI custom nodes can execute code and make network requests independently of the model.
• Remote-access software can expose the interface or capture the screen.
• Observability and crash-reporting systems can record request bodies or file paths.
• Cloud-synced output folders can upload WAV files after generation.

The privacy claim should therefore be phrased precisely: Music 3 supports a local inference architecture. Whether your lyrics stay local is a property of your complete installation and operating procedure.

The 57 GB Download And The 8 GB VRAM Claim Are Both True

Disk storage and GPU memory answer different questions.

The 57.35 GB repository total describes the artifacts stored on disk. VRAM describes how much of the active pipeline must reside on the graphics card at a given time. CPU offloading or layer streaming can keep some components in system memory and move them to the GPU when needed.

That trade can reduce peak VRAM without shrinking the model files. It can also increase load time and generation time because data moves across a slower path.

Budget for more than the headline download:

• The full repository and temporary download space.
• Python environments, CUDA libraries, and serving-framework dependencies.
• System RAM for CPU offload or streaming.
• Output WAV files and alternate generations.
• Cache directories that may duplicate downloaded files.
• Workspace copies created by ComfyUI, notebooks, or scripts.
• Backups, snapshots, and rollback copies.

A five-minute stereo WAV at 32 kHz and 16 bits is not enormous by itself, but repeated seeds, revisions, stems created by other tools, and project backups can accumulate quickly. More importantly, every additional copy becomes another place where deletion and access controls must work.

Use The SONG Audit Before Trusting Unreleased Material

The SONG audit separates local model execution from the surrounding creator workflow.

S — Source And Staging

Record where the model, code, dependencies, lyrics, references, and prompt templates come from. Pin repository revisions when possible and review unexpected executable files or custom nodes before use.

Download acquisition is not the same as generation. Hugging Face and package registries necessarily see the download requests made to them. After staging, test whether generation still works with outbound network access blocked.

Do not begin with a valuable lyric. Use a synthetic verse containing a unique, harmless canary phrase that appears nowhere else. If that phrase later appears in an unexpected log, cache, request capture, or synced folder, you have found a copy path without exposing real work.

O — Outbound Connections

Monitor DNS queries and network connections while launching the interface, loading the model, enhancing the prompt, generating audio, previewing it, and saving or exporting the result.

Run the same workflow with outbound access denied. Local generation continuing successfully is strong evidence that the active inference path does not require a remote service. Then enable optional features one at a time and observe what changes.

A blocked connection is not automatically data leakage. It may be an update check that sends no prompt content. Investigate the destination, method, request body, and timing before drawing a conclusion.

N — Notes, Names, And Networked Tools

Inventory every input and identifier that enters the workflow: lyrics, artist names, collaborator names, voice descriptions, production notes, file paths, seed values, and reference audio used by surrounding tools.

Then map connected tools. Prompt rewriters, web search, cloud drives, source control, transcription, voice conversion, mastering services, and collaboration platforms can each send data elsewhere even when Music 3 itself runs locally.

Use the narrowest possible data. A style description usually does not need a collaborator's legal name, email, contract terms, home address, or unreleased business plan.

G — Generated Files And Garbage Collection

Find every durable copy of inputs and outputs. Check the project directory, model cache, UI history, temporary folders, shell history, notebooks, API logs, browser storage, recent-file lists, thumbnails, backups, and cloud sync.

Define deletion before production use. Deleting the final WAV is incomplete if the prompt remains in a UI database, a notebook cell, an API log, or a backup. Conversely, clearing a UI history does not prove that the WAV disappeared from disk.

Repeat the canary test after deletion. Search approved storage for the unique phrase and verify that intended retention and backup rules behave as expected.

A Practical Private Music 3 Setup

A defensible local workflow can be built in stages:

1. Use a dedicated operating-system account or isolated machine for sensitive projects.
2. Download the official repository and required packages before importing unreleased material.
3. Pin known versions of the runtime and review third-party wrappers or custom nodes.
4. Bind the generation service to loopback unless another device genuinely needs access.
5. Require authentication and a firewall if the service must listen beyond loopback.
6. Disable optional analytics, cloud storage, hosted prompt enhancement, and remote tools.
7. Generate once with synthetic canary lyrics while monitoring network and filesystem activity.
8. Block outbound access and repeat the full generation and export path.
9. Store lyrics and outputs in an access-controlled location outside automatic cloud sync.
10. Document how prompt records, caches, outputs, backups, and temporary files are deleted.

Containerization can improve repeatability, but a container is not automatically a security boundary. Mounted folders, host networking, environment variables, GPU access, and privileged execution still matter. A local web server exposed to the LAN can also become a data path if another device or user can reach it.

For an especially sensitive project, keep the generation computer disconnected after all artifacts are staged, and move only approved outputs through a controlled process. That is more operationally demanding, but it is a clearer boundary than relying on an unchecked “offline” toggle.

What The Music 3 License Changes

MiniMax Music 3 uses a custom Community License, not Apache 2.0 or MIT for the entire release.

The license broadly permits use, modification, distribution, and hosted services, subject to conditions. Among other terms, it requires a commercial product or service using the software to display “MiniMax-Music3” prominently in its interface. It requires separate prior written authorization when aggregate yearly revenue generated by covered products or services from the user and affiliates exceeds $20 million.

If a product lets third parties generate outputs with Music 3, the operator must implement and periodically review reasonable safeguards against prohibited or rights-violating access, uses, and outputs. The acceptable-use policy also requires public machine-generated material to be clearly disclosed as machine-generated and prohibits a range of harmful or unlawful uses.

Those are operational obligations, not privacy guarantees. A license can require safeguards without proving how a particular deployment handles lyrics, logs, or generated audio. Commercial teams should review the full text for their use case rather than relying on a model-card badge or a generic “open source” description.

Local Music Generation Versus A Hosted Service

Local Music 3 is the stronger choice when active lyrics and audio must remain on hardware you control, you have compatible CUDA hardware, and you are willing to operate and audit the stack.

A hosted music service may be easier when generation speed, mobile access, collaboration, or avoiding a 57 GB model matters more. But active prompts and outputs necessarily have to be processed on that service's infrastructure under its current policies and account settings.

The choice is not “local is always private, hosted is never private.” It is a choice between trust boundaries:

• Local execution gives you more control and more responsibility for patches, network policy, logs, users, storage, backups, and deletion.
• Hosted execution delegates infrastructure while adding provider processing, account, retention, and access questions.
• A hybrid workflow can keep generation local while still leaking information through cloud notes, prompt tools, collaboration, or output sync.

Our guide to private AI versus local AI explains the broader tradeoff. Also remember that local AI can use web search and still send queries outward; the same principle applies to connected creator tools.

Where OpenVeil Fits—And Where It Does Not

OpenVeil is a hosted, privacy-focused AI workspace for adults who want chat, private search, files, voice, images, video, and custom personas without assembling a local model stack.

Normal OpenVeil chat history stays in the user's browser, and OpenVeil does not maintain a normal server-side chat-history record for private chat sessions. OpenVeil does not use prompts, uploaded files, images, audio, selected local-history context, or AI outputs to train foundation models.

OpenVeil is not a music generator or a MiniMax Music 3 host. It is not fully offline, anonymous, zero-log, or a substitute for self-controlled inference when active lyrics must never leave hardware you control. Active requests still require processing by OpenVeil and necessary providers, including AI, search, upload-processing, hosting, routing, security, billing, and infrastructure providers.

The natural split for a creator is to use Music 3 locally for song generation when that execution boundary matters, and use a separate privacy-focused workspace for research, planning, documents, conversation, or media tasks that fit its documented capabilities. If a hosted workspace with browser-local normal history and no foundation-model training on your prompts or files fits that boundary, try OpenVeil.

Before trusting either option, check each AI privacy claim against active processing, history, retention, training, network tools, deletion, and provider access.

Frequently Asked Questions

Can MiniMax Music 3 run completely offline?

It can support offline generation after the model files, runtime, dependencies, and CUDA environment are already present. The official examples do not require a hosted MiniMax generation call. You still need to block and monitor outbound access to verify your exact runtime, UI, plugins, and export workflow.

Does MiniMax Music 3 send lyrics to MiniMax?

Not inherently when you use the downloadable weights with a local runtime. The official SGLang example sends the request to a loopback address on the same machine. A third-party wrapper, prompt tool, analytics service, remote interface, or cloud-synced folder can create separate outbound paths.

How large is the MiniMax Music 3 download?

The current Hugging Face repository lists 88 files totaling 57,353,379,600 bytes, about 57.35 GB or 53.41 GiB. Allow additional space for temporary downloads, Python and CUDA dependencies, caches, outputs, and project copies.

Can Music 3 run on an 8 GB GPU?

MiniMax's model card says layer-by-layer language-model streaming can make the pipeline fit on an 8 GB GPU. That path is described as slower. CUDA is required, and system RAM, storage, runtime compatibility, and acceptable generation speed still need to be validated on the actual computer.

Does local Music 3 guarantee that no logs exist?

No. The serving framework, UI, operating system, command history, notebooks, monitoring, remote-access software, and backups can create records. Local execution gives the operator control over those systems; it does not configure them automatically.

Is MiniMax Music 3 open source?

“Open-weight release under the MiniMax-Music3 Community License” is the more precise description. The weights and code are downloadable, but the custom license contains attribution, revenue-threshold, safety, disclosure, and acceptable-use conditions.

Is OpenVeil a private alternative for generating songs?

No. OpenVeil does not document music generation. It is a hosted privacy-focused workspace for chat, files, search, voice, images, video, personas, and related planning or analysis. Active requests are processed by OpenVeil and necessary providers, so it is not an offline replacement for local Music 3.

The Bottom Line

MiniMax Music 3 makes private local song generation technically credible: the weights are downloadable, the official serving paths include loopback and in-process generation, and the model card documents ways to fit the pipeline from under 24 GB down to an 8 GB GPU with performance tradeoffs.

What it does not provide is a universal privacy guarantee. The 57.35 GB acquisition path, dependencies, UI, plugins, logs, remote access, backups, and connected creator tools all sit outside the model's core inference claim.

Use the SONG audit with synthetic lyrics, verify generation while outbound access is blocked, find every stored copy, and test deletion before importing unreleased work. “Local” is the beginning of the privacy review—not the conclusion.