Product and research teams building song generation into an app, rather than solo creators wanting a browser studio, are the natural fit for MiniMax Music 3.0 against closed 2026 rivals like Suno and Udio. It ships open weights, splits reference inference across 2 CUDA GPUs, and MiniMax has not published a single competitive benchmark score.
MiniMax Music 3.0 is an open-weight text-to-music model that MiniMax released in 2026, pairing a large language model with a flow-matching module to generate complete 5-minute songs, full vocals and arrangement included, from one text prompt plus optional section-tagged lyrics.
Provider: MiniMax · Family: MiniMax Music
Input modalities: text · Output: audio
About MiniMax Music 3.0
MiniMax Music 3.0 is a text-to-music generation model built by MiniMax, the Shanghai AI lab that also builds the MiniMax M-series language models and Hailuo video models. MiniMax released Music 3.0 through its hosted API on July 16, 2026, then followed up four weeks later with a Hugging Face weight release under a new MiniMax-Music3 Community License. Music 3.0 replaces Music 2.6 as MiniMax's flagship music model and is built to take one creative brief, a style and mood description plus optional lyrics, and return a finished song rather than a short loop or instrumental sketch. The model is a four-stage hybrid system: an 8 billion parameter Global LLM (initialized from Qwen3.5-8B) predicts semantic tokens frame by frame for structure, a 0.6 billion parameter Local LLM (trained from scratch) fills in frame-level acoustic detail, a 2.4 billion parameter flow-matching module fuses both hidden states, and a 123 million parameter Flow-VAE decoder renders the waveform, for roughly 11.1 billion parameters combined. Training used an 8-layer RVQ tokenizer: one 16,384-entry codebook for musical semantics and seven 1,024-entry codebooks for acoustic residue. A single generation produces up to five minutes of finished song at 32 kHz, 16-bit stereo WAV. MiniMax designed the model to sustain one creative intent, a mood, a story, a hook, across the full song instead of losing coherence after the first chorus, and to keep vocal identity, key, and arrangement consistent from intro to outro. There is no context window in the language-model sense; length is capped by the output ceiling and by a default 750-frame generation chunk that the reference implementation can extend up to 9,000 frames per call. Input is a text prompt describing genre, tempo, key, and production style, plus optional lyrics written with section tags: Intro, Verse, Pre-Chorus, Chorus, Post-Chorus, Bridge, Instrumental, Solo, and Outro. A lyrics_optimizer parameter writes full lyrics from a one-line theme for users with none of their own. A separate is_instrumental flag skips vocals entirely, and vocal controls choose male, female, or duet delivery. MiniMax also ships a companion audio-to-audio model, Music Cover, which takes an uploaded song, preserves its melody, and re-renders it in a new genre, tempo, and vocal timbre. Neither model does image input, web browsing, or general tool calling; the entire interface is the structured text prompt. MiniMax bills Music 3.0 per generation rather than per token, with a metered free tier and a higher-throughput paid tier; the lyrics_optimizer add-on and the companion Music Cover model bill the same way. Exact current rates are in the pricing FAQ below and in price_notes, since per-generation prices for hosted music models change often. Third-party hosts such as fal.ai and WaveSpeedAI mirror the model behind their own endpoints and set independent markups on top of MiniMax's base price. The hosted API sits at api.minimax.io/v1/music_generation, authenticated with a bearer token and called with the model parameter set to music-3.0. Because the weights are open, teams can also self-host: the Hugging Face repository, MiniMaxAI/MiniMax-Music3, is about 57.4 GB in Safetensors format. The reference setup splits inference across two CUDA GPUs, one running the Qwen3-based language model and RVQ token generation, the other running flow matching and waveform decoding. On a single GPU, MiniMax states 24 GB or more VRAM for full precision, with CPU offloading or layer-by-layer streaming dropping that further at reduced speed. Community GGUF, INT4, and NVFP4 quantizations are already on Hugging Face for lower-memory consumer GPUs. MiniMax has not published a system card, red-team disclosure, or training-data source list for Music 3.0, and neither the API docs nor the GitHub repository state whether copyrighted recordings were used in training. The open-weight release carries no audio watermark, a real difference from Google's Lyria, which marks every output with SynthID under an explicit no-licensed-training claim. MiniMax's separate browser-based consumer product requires labeling AI-generated output and bars removing any required watermark, but that policy sits on the consumer app's terms of service, not on the open weights or the raw API. Music 3.0 fits teams that want a repeatable API for full songs inside a product, an app that needs melody-preserving covers, or a research group that wants to fine-tune or run a music model on its own hardware instead of paying per track. It is a weaker fit for solo creators who want a browser tool with a polished mixing and download workflow, since third-party reviews consistently place two closed rivals ahead of it on vocal appeal and final mix quality. Teams needing contractual proof of licensed training data or automatic content labeling should look elsewhere until MiniMax publishes that documentation. Music 3.0 is the fifth release in MiniMax's music line, after Music 2.0 (late 2025, emotional expressiveness), Music 2.5 (January, finer detail control), Music 2.5+ (March, pure instrumental generation), and Music 2.6 (April, rebuilt cover feature and low-end bass). Its advance over that line is the move to the four-stage hybrid architecture described above and to open-weights distribution rather than an API-only model.
Pricing
Paid-tier rate: $0.15 for one finished song of up to five minutes, at up to 120 requests per minute; add $0.01 per track if lyrics_optimizer writes the lyrics. Music Cover (audio-to-audio style transfer) bills the same $0.15 per generation. Free tier: capped at 3 requests per minute, no charge. Verified against MiniMax's live pay-as-you-go pricing docs, August 2026.
Key Features
- Four-stage hybrid architecture: An 8B Global LLM (from Qwen3.5-8B) holds song structure, a 0.6B Local LLM fills in acoustic detail, a 2.4B flow-matching module fuses both, and a 123M Flow-VAE renders the final waveform.
- Full 5-minute song generation: Composes, arranges, performs, and produces a complete song up to five minutes in one generation call, holding theme, key, and vocal identity from intro to outro.
- Structured lyrics and caption input: Accepts lyrics tagged by section (Intro, Verse, Chorus, Bridge, Outro, and more) plus a structured caption covering genre, BPM, key, vocal timbre, and instrument-by-section arrangement.
- Lyrics optimizer: Set lyrics_optimizer to true and the model turns a one-line theme into a complete lyric sheet for users who don't bring their own, billed as a small per-song add-on.
- Open weights on Hugging Face: MiniMaxAI/MiniMax-Music3 ships as a Safetensors repo (~57.4 GB) under a free community license, deployable via ComfyUI, Diffusers, or SGLang-Omni nodes.
- Music Cover audio-to-audio model: A companion model, music-cover, uploads an existing song and reimagines its genre, tempo, and vocal timbre while keeping the source melody intact.
Pros
- Generates a complete five-minute song, vocals and full arrangement included, from a single API call instead of a short loop or clip.
- MiniMax's published per-generation rate is the lowest among frontier music vendors as of 2026; see the pricing FAQ for the exact figure.
- Ships as open weights for free under a license that only requires separate authorization above $20 million in annual revenue.
- The companion Music Cover model reuses the same API shape and pricing for audio-to-audio style transfer, so teams don't need a second vendor for covers.
Cons
- No published benchmark or blind-listening score against Suno, Udio, or Google Lyria, so quality claims rest on third-party reviews, not MiniMax's own data.
- Multilingual lyrics are noticeably weaker than English in independent testing, particularly Hindi and Spanish.
- Self-hosting the open weights needs 24 GB or more of VRAM for full precision, so it isn't practical on most consumer laptops without quantization.
Frequently Asked Questions
How much does MiniMax Music 3.0 cost per song in 2026?
MiniMax charges $0.15 per generation for up to five minutes of finished audio on the paid tier, plus $0.01 per song if the lyrics_optimizer writes the lyrics for you. The paid tier allows 120 requests per minute, while a free tier caps out at 3 requests per minute with no cost. Third-party hosts such as fal.ai and WaveSpeedAI set their own markups on top of this base price.
How does MiniMax Music 3.0 compare on quality to Suno and Udio?
MiniMax has not published a benchmark or blind-listening score for Music 3.0 against any competitor. Independent reviewers generally rate Suno ahead on vocal catchiness and Udio ahead on overall mix polish, while crediting MiniMax with strong raw audio fidelity and the lowest per-track price of the three as of 2026.
Is MiniMax Music 3.0 open source?
MiniMax released the Music 3.0 weights on Hugging Face on August 13, 2026 under the MiniMax-Music3 Community License, which lets anyone use, modify, and distribute the model for free. Companies with aggregate yearly revenue above $20 million need separate written authorization from MiniMax before commercial deployment. Unlike MiniMax's H3 language model license, this one carries no territorial restrictions.
Does MiniMax Music 3.0 train on user data or watermark its output?
MiniMax has not published a data retention or training-on-inputs policy for the Music API as of 2026. Neither the hosted API nor the open weights add an audio watermark to generated tracks, unlike Google's Lyria, which watermarks every output with SynthID. MiniMax's separate consumer browser app requires labeling AI-generated content, but that rule does not extend to API or self-hosted output.
Who should use MiniMax Music 3.0, and who should avoid it?
Music 3.0 fits product teams that want a low-cost, repeatable song-generation API or a self-hostable open model for research and fine-tuning. Solo creators who want a polished browser studio with stems and easy editing are generally better served by Suno or Udio. Teams that need certified training-data provenance or automatic content labeling should wait for MiniMax to publish that documentation.