MiniMax Music 3: Running Locally, License, Pricing and Quality

👁 116 reads

MiniMax Music is an AI song generator created by Shanghai-based MiniMax, the team behind Hailuo Video. Its latest release, Music 3, creates full vocal tracks up to 5 minutes long and has offered open weights since August 2026, meaning you can run it for free on your own GPU via ComfyUI. It sings clearly across languages though accented in non-native tongues, and its paid direct API closed to new customers on August 20.

What is MiniMax Music?

MiniMax is a Shanghai AI company known for large language models, video generators (Hailuo, H3), and voice synthesis. I cover their video ecosystem in the MiniMax: Hailuo, H3 and H3 Max guide; here I focus strictly on their music generation.

MiniMax Music operates similarly to Suno: you provide a style description and lyrics, and the model generates the melody, instrumental arrangement, and vocal performance. Music 3 accepts two primary inputs:

• Structured lyrics - section tags such as [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo], and [Outro]. The model follows these markers closely.
• Structured Caption - a prompt divided into three components: Global Metadata (genre, tempo, key, mood), Vocal Details (vocal type, delivery style), and Arrangement (instrumentation and structural changes across sections).

Under the hood, an 8B Global LLM (based on Qwen3-8B) maintains macro-structure, a 0.6B Local LLM handles granular transitions, a 2.4B Flow Matching module synthesizes sound, and a 123M Flow-VAE outputs 32 kHz stereo WAV files.

MiniMax Music version history

Release dates come directly from MiniMax release notes (2.5+ and 2.6 were announced by month only). The Hugging Face repository was initiated on August 7, followed by the public launch from MiniMax and ComfyUI on August 13, 2026.

Benchmark rankings: how MiniMax Music stacks up

Artificial Analysis version 1.1 of its Text-to-Music benchmark (around October 7, 2026, covering 1,000 prompts across 17 genres) calibrated Music 3.0 as its baseline reference score of exactly 1,000. All comparative data is cataloged in my AI music generator ranking.

• Vocal tracks: Music 2.6 scored 1,001 (10th place), Music 3.0 scored 1,000 (11th place), Music 2.5+ scored 987 (12th place). The overall leader, Suno v6, achieved 1,142.
• Instrumentals: Music 2.5+ reached 1,005 (10th place), Music 2.6 hit 1,002 (11th place), and Music 3.0 scored 1,000 (13th place). Suno v6 leads here with 1,140.
• Open weights: Music 3.0 is the highest-rated downloadable model you can host entirely on your own hardware.

Music 3 does not sound noticeably superior to Music 2.6 in blind quality tests; its real advantage is local self-hosting rather than a leap in raw output fidelity.

MiniMax-Music3 open-weights license

The released weights do not use Apache 2.0 or MIT, but instead use a proprietary “MiniMax-Music3 COMMUNITY LICENSE”. Checking the repository LICENSE file (as of October 8, 2026), the core terms include:

• Commercial use is permitted, including embedding the model into paid products or hosted cloud platforms.
• Attribution requirement: Any commercial service powered by the model must visibly display “MiniMax-Music3” in its interface.
• $20 million revenue ceiling: If gross annual revenues from products or services utilizing the model exceed $20 million, you must request prior written permission from MiniMax (via api@minimax.io).
• Safeguards: Providers hosting the model for third parties must implement reasonable, proportionate safeguards to prevent prohibited outputs.
• Acceptable Use Policy: An attached policy (dated August 6, 2026) bans military applications, disinformation campaigns, impersonation, and harmful content.

Attribution requirements and revenue caps apply to software products and services hosting the weights, not to the rendered music tracks. The license does not claim ownership over the generated audio output. A song produced locally on your rig and published on streaming services does not require a “MiniMax-Music3” badge in the title, though you remain subject to the Acceptable Use Policy. References to CC BY-NC 4.0 found on some discussion boards are incorrect.

MiniMax Music pricing

MiniMax Audio web app

Web generations are created at MiniMax Audio (minimax.io/audio/music), formerly hailuo.ai/audio. Credit pools are shared across music and speech synthesis. One song generation consumes 300 credits (generating two parallel variations costs 600).

Pricing verified at minimax.io/audio/subscribe on October 8, 2026. Commercial rights to newly generated songs are granted under paid tiers. For the Free plan, site copy is contradictory: tier cards suggest commercial usage, while the platform FAQ limits free tiers to personal evaluation. Follow the FAQ guidance. The platform terms of service (dated June 19, 2025) require disclosing AI involvement and prohibit removing platform audio watermarks.

API access

• MiniMax Native API: Music-3.0 and Music-2.6 were priced at $0.15 per track up to 5 minutes, with lyrics generation billed at $0.01. Since August 20, 2026, new account signups are blocked from purchasing API credits; existing funded accounts continue operating.
• fal.ai: minimax/music-3 costs $0.002 per second of generated audio ($0.12 per minute). Music 2.6 (fal-ai/minimax-music/v2.6) costs a flat $0.15 per track, based on fal pricing API records from October 8, 2026.

How to run MiniMax Music 3 locally in ComfyUI

Running via ComfyUI is the most accessible path. MiniMax also outlines setups using diffusers (their documentation notes 24 GB VRAM baseline, ~22 GB with CPU offload, and down to 8 GB using layer streaming) alongside SGLang-Omni. An NVIDIA GPU with CUDA support is mandatory. For wider GPU planning, consult the guide on how much VRAM you need, and see my setup walkthrough for ComfyUI.

1. Update ComfyUI to the latest release to ensure MiniMax Music 3 audio nodes are recognized.
2. Open the Template Library, enter the Audio section, and load the audio_minimax_music_3 workflow (“Text to Music (MiniMax Music 3)”). ComfyUI will prompt you for missing weights.
3. Download the model files from the Comfy-Org/MiniMax-Music-3 repository into your installation directories:
   • models/diffusion_models/: minimax_music3_dit_fp16.safetensors (4.91 GB), or for tighter budgets, minimax_music3_dit_int8_convrot.safetensors (2.5 GB),
   • models/text_encoders/: minimax_music3_text_encoder_pruned_int8_convrot.safetensors (9.2 GB),
   • models/vae/: minimax_music3_dav.safetensors (0.22 GB).
4. Inside the MiniMaxMusic3TextEncode node, supply your three-part prompt (Global Metadata, Vocal Details, Arrangement) and lyrics separated by structural tags.
5. In the EmptyMiniMaxMusic3LatentAudio node, define output duration in seconds. The default preset is 120s, with a maximum around 300s.
6. Keep standard KSampler workflow defaults: 30 steps, cfg 1.7, euler sampler, simple scheduler.
7. Audio decoding: use VAEDecodeAudio for high-VRAM setups, or toggle VAEDecodeAudioTiled (via the tiled_decode switch) if running low on memory during long generations. Tiled decoding runs slower and can occasionally cause minor transition artifacts.
8. Queue the prompt. Finished audio outputs as MP3 files in output/audio/.

What are the hardware limits? The text encoder is considerably heavier than the audio diffusion model itself. Loading the int8 package (DiT 2.5 GB + text encoder 9.2 GB + VAE) takes roughly 11.9 GB of storage; the fp16 DiT stack requires ~14.3 GB. Community-quantized GGUF versions include:

• Abiray/MiniMax-Music3-GGUF - quantizes the DiT component: spanning Q3_K_M (1.16 GB) up to Q8_0 (2.7 GB) and F16 (4.98 GB). The encoder and VAE still load from Comfy-Org.
• molbal/Minimax-Music3-GGUF - quantizes both the DiT and text encoder, shrinking the encoder to roughly 5.4 GB in Q4_0. This requires a compatible ComfyUI-GGUF fork.

According to localaimaster.com, a practical baseline configuration is a 12 GB GPU paired with 32 GB of system RAM. An 8 GB card functions only with heavy offloading layers and runs slowly. For further open-weights alternatives, visit the Local AI index.

Hardware benchmark: MiniMax Music 3 on an RTX 4090

On October 8, 2026, I tested MiniMax Music 3 locally on an RTX 4090 via ComfyUI (fp16 DiT at 4.91 GB, pruned int8 text encoder at 9.2 GB, and VAE at 0.22 GB, totaling roughly 14.3 GB downloaded from Comfy-Org on Hugging Face). Pipeline parameters matched the default workflow: 30 steps, cfg 1.7, euler/simple, capped at 100 seconds. We tested it in Polish using an original test track titled “Morning in Lodz” (translated from the Polish original) (2 verses and a chorus, 52 words) alongside a style prompt specifying a nostalgic indie pop ballad with female vocals.

• Vocal articulation: Whisper large-v3, applied to stems extracted via Demucs, correctly transcribed 102 out of 104 words across both runs. The two phonetic mistakes involved Polish soft-consonant inflections: singing “wyszle” instead of “wysle” and “pognieciona” with an exaggerated palatalized vowel.
• Stylistic fidelity: Automated evaluation models (Gemini 3.8 Flash and Qwen 3.8 Omni Flash) scored pronunciation at an average of 4.0/5, but assigned only 3.1 for timbre and 2.9 for genre adherence. Gemini reported that rather than producing an intimate indie ballad, the generator delivered an overly theatrical, Broadway-style vocal with noticeable metallic artifacts.
• Duration variance: Track lengths proved irregular: the second attempt terminated at 54.5 seconds, delivering all lyrics at a rushed tempo rather than sustaining the requested pacing.
• Comparative baseline: ACE-Step 1.5 tested on the same RTX 4090 rendered 95 seconds of audio in 8 seconds flat while drawing 13-14 GB of VRAM, scoring 4.5 on phonetic clarity.
• Cloud hosted run: In an earlier test on October 3 via fal.ai, MiniMax Music 3 scored worse on a different non-English prompt, getting 32 of 39 words right with a 3/5 rating.

Complete multi-model results are documented in the AI music generator ranking. Initial run times were lengthened by loading weights from an external USB drive (~44 MB/s); performance will improve significantly on fast internal NVMe storage.

Multilingual performance and vocal clarity

MiniMax does not publish a formal list of supported vocal languages. Official Music 3 demos focus strictly on Mandarin Chinese and English. Marketing copy citing “40+ languages” across MiniMax Audio refers solely to its text-to-speech platform, not vocal singing.

I investigated cross-language fidelity on October 3, 2026, for the AI music generator guide using identical non-English test lyrics (we tested it in Polish, 39 words), transcribing the audio with Whisper large-v3 and rating vocal delivery via Gemini.

• Music 2.6: 34 out of 39 words recognized, pronunciation rated 2/5, 57-second generation latency on fal.ai.
• Music 3: 32 out of 39 words recognized, pronunciation rated 3/5, 120-second generation latency. Output duration was shorter than prompted (27 seconds instead of the requested 75 seconds).
• Common failures: Both models struggled with dense consonant clusters (such as “z Gdanska rusza”), whereas melodic open-vowel choruses rendered cleanly.

Best practice: write the descriptive style prompt in English while adding clear instructions like “vocal singing with clear diction”, while keeping the raw lyrical passages in your target language without overly dense consonant openings in the first bar.

MiniMax Music vs competing AI generators

The highest foreign-language articulation score in my testing was achieved by Eleven Music (39/39 on October 3, and 103/104 in follow-ups). Among self-hosted models, Music 3 ranks highest in the Artificial Analysis leaderboard; however, in my local tests on October 8, ACE-Step 1.5 achieved comparable phonetic clarity (101/104 words), stayed closer to the requested musical style, and required less GPU memory. An earlier version of ACE-Step on fal.ai had broken down into vocal distortion on non-English inputs (8/39 words on October 3), but version 1.5 is a substantially redesigned model.

Who should use MiniMax Music 3?

• Hardware owners with 12-24 GB VRAM wanting zero-subscription music generation and unlimited local audio exports.
• Software developers seeking a robust open-weights audio engine with commercial deployment rights (subject to the “MiniMax-Music3” naming attribution).
• English vocal and instrumental production, where musical styling, dynamic range, and tonal balance are strongest.
• Not ideal for commercial regional-language vocals without post-tuning, due to audible accenting. Dedicated vocal tools like Eleven Music or Lyria deliver cleaner native diction.

Frequently asked questions

What is MiniMax Music?

MiniMax Music is an AI-driven song generation engine built by Chinese AI developer MiniMax. Given a text prompt and lyrics, it produces fully arranged tracks up to 5 minutes long. Version 3 features downloadable open weights and can be run locally or accessed through the MiniMax Audio platform and fal.ai.

Can I run MiniMax Music 3 locally on my computer?

Yes. The simplest method is via ComfyUI using the official audio_minimax_music_3 workflow template. You will need an NVIDIA graphics card with at least 12 GB VRAM for comfortable inference, and roughly 12 to 15 GB of drive storage for model checkpoints. Quantized GGUF builds can operate on 8 GB cards when utilizing aggressive CPU offloading, albeit much slower.

Is MiniMax Music free to use?

The open-weights model is completely free to self-host. In the cloud web app (MiniMax Audio), free tiers receive 10,000 monthly credits (around 33 songs), but the platform FAQ reserves free outputs for personal non-commercial testing. The free public API endpoints were discontinued on August 20, 2026.

What license governs MiniMax Music 3 weights?

Weights are distributed under the MiniMax-Music3 Community License. Commercial product integration is permitted, provided your app clearly shows “MiniMax-Music3” attribution and your company generates under $20 million in annual gross revenue. These restrictions apply to software products and services hosting the model, not individual exported songs.

Does MiniMax Music sing in languages other than English and Chinese?

Yes, though pronunciation accuracy varies and accents are present. In benchmark runs, Music 3 rendered 32 out of 39 non-English words accurately in early cloud tests, and up to 102 out of 104 words during local inference. Official model documentation focuses primarily on English and Chinese.

What is Hailuo Music?

Hailuo Music was the original branding for MiniMax’s audio generation tool within Hailuo AI. The music service has since been unified under MiniMax Audio at minimax.io/audio/music. For an evaluation of their video system, see my breakdown of Hailuo AI.

Can I monetize tracks generated with MiniMax Music?

Songs generated under paid MiniMax Audio subscription tiers carry commercial rights, provided you adhere to platform guidelines regarding AI content labeling. When running the open weights locally, the community license does not claim ownership over generated audio files or prohibit selling rendered music.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and verification dates

• Model card and LICENSE file for MiniMaxAI/MiniMax-Music3 - huggingface.co (verified October 8, 2026)
• Comfy-Org/MiniMax-Music-3 release files - huggingface.co (October 8, 2026)
• MiniMax Music 3.0 launch post - minimax.io (August 13, 2026)
• Platform model release documentation - platform.minimax.io (October 8, 2026)
• API pricing structure and deprecation advisory - platform.minimax.io (October 8, 2026)
• MiniMax Audio subscription tiers - minimax.io (October 8, 2026)
• MiniMax Audio terms of service - minimax.io (dated June 19, 2025)
• ComfyUI blog coverage on Music 3 - blog.comfy.org (August 13, 2026)
• GGUF repositories - Abiray and molbal (October 8, 2026)
• Local deployment VRAM profiles - localaimaster.com (October 8, 2026)
• Artificial Analysis Text to Music benchmark v1.1 - Artificial Analysis (approx. October 7, 2026)
• fal.ai pricing API records (October 8, 2026) and test generation run (October 3, 2026)

Kinetyra