# MiniMax Music 3 Review: Is This Open AI Music Model Any Good? > Source: > Published: 2026-08-14 00:00:00+00:00 # MiniMax Music 3 Review: Is This Open AI Music Model Any Good? A hands-on test of MiniMax Music 3 across pop, Bollywood, and cumbia genres finds solid English vocals but weak multilingual output. ## What is MiniMax Music 3? MiniMax Music 3 is an open music generation model that turns lyrics and a text description into a complete song, up to five minutes long, with vocals, instrumentation, and full arrangement. Feed it a genre, tempo, key, and vocal style, and it writes lyrics on its own if you don’t supply them, then renders a structured track with intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro. Output comes as 32 kHz, 16-bit stereo WAV. The pitch is straightforward: give the model a prompt, get a finished song, not just a loop or a beat. ## TL;DR - MiniMax Music 3 generates **full structured songs**, including intro, verses, chorus, bridge, and outro, from a text prompt and optional lyrics, outputting 32 kHz 16-bit stereo WAV. - The model uses a **hierarchical autoregressive architecture**: an 8-billion-parameter global LLM (initialized from a 1.3B base) handles long-range structure, while a smaller 0.6B local LLM fills in fine acoustic detail within each frame. - It relies on an **eight-layer RVQ tokenizer** and flow matching, plus a flow variational autoencoder adapted from MiniMax’s speech model, to convert latent representations into final audio. - In testing, an **acoustic pop track in English** came out coherent, with structurally sound arrangement and generally listenable vocals. - A **Bollywood-style item song attempt** produced vocals and instrumentation that bore little resemblance to actual Hindi film music, with pronunciation and melodic phrasing well off the mark. - A **cumbia generation attempt** also missed the genre’s actual rhythmic and instrumental character, despite the prompt specifying accordion, brass, and dance rhythm. - Running locally required **around 27 GB of VRAM** on an RTX 6000 (48 GB card), tested via a Gradio interface built on top of MiniMax’s release script. ## Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. ## How do you run MiniMax Music 3 locally? The model can be installed on a Linux machine with a capable Nvidia GPU. The setup shown in testing used an Ubuntu system with an RTX 6000 card carrying 48 GB of VRAM, though actual consumption during generation measured just over 27 GB, meaning it should run on smaller cards too, depending on how much headroom you want. The install path is short: set up a Python environment, install the `diffusers` library and `torchaudio` , then run the generation script MiniMax provides. Wrapping that script in a Gradio interface gives you a browser-based UI where you can type a prompt, hit generate, and wait for the model to write lyrics, build a structured caption, and render the audio. On first run, the script also handles downloading model weights. There are two main ways to interact with it. A simple mode lets you type a rough description (genre, key, mood) and let the model handle the rest, including writing its own lyrics. A studio mode gives more granular control: you write the lyrics yourself, specify section-by-section arrangement, and set vocal details like gender, tone (breathy, tenor, etc.), instrumentation, and even a title. For anyone wanting reproducible output rather than a lucky roll, studio mode is the more useful entry point. ## How does MiniMax Music 3 actually work under the hood? The architecture splits the generation task into two cooperating language models. A larger global LLM, roughly 8 billion parameters and initialized from a 1.3-billion-parameter base, predicts the first RVQ (residual vector quantization) codebook frame by frame. This model is responsible for the song’s overall structure and long-range musical coherence, essentially deciding how the track evolves from intro to outro. A smaller local LLM, around 0.6 billion parameters, then fills in the remaining acoustic codebooks within each frame, restoring the fine-grained texture that makes vocals and instruments sound natural rather than flat. Rather than decoding purely from discrete tokens, the system fuses continuous hidden states from both models before handing off to synthesis, which is meant to preserve more detail for vocal articulation and instrumental texture. Underneath this sits an eight-layer RVQ tokenizer used during training, along with flow matching, a generative technique that learns to transform random noise into target audio by following a smooth, continuous path rather than jumping directly to a finished waveform. At the final stage, a flow variational autoencoder, adapted from MiniMax’s speech model and retrained specifically for music, decodes the latent representation into the finished stereo waveform. It’s a layered pipeline: structure and semantics from the large model, acoustic detail from the small model, and waveform synthesis from a dedicated decoder. That division of labor is common in modern audio generation, where one network plans and another one paints in detail. ## Is MiniMax Music 3 good at generating English pop music? ## Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. Based on hands-on testing, the answer leans toward yes for straightforward English content. A prompt specifying acoustic pop, a warm and building arrangement, and a key of C major produced a coherent song: the model wrote its own lyrics, built a full arrangement, and delivered vocals that held together across the track. The audio was clean, structurally sound, and didn’t fall apart over the run time, which is a real technical achievement given how many generation models struggle to maintain coherence past 30 or 60 seconds. That said, judging vocal quality is subjective, and the reviewer running the test noted having limited familiarity with Western music production standards, so listeners with more genre expertise may hear things a casual listener would miss. Still, on its own terms, the English pop result did what it promised: readable structure, stable audio, and vocals that sounded like a song rather than a demo. ## How well does it handle non-English and non-Western music? This is where the cracks show. MiniMax Music 3’s model card doesn’t make explicit multilingual claims, and testing suggests that gap matters. A Bollywood-style item song prompt, complete with custom lyrics and vocal metadata specified through studio mode, produced results described as far off the mark compared to the actual sound of Hindi film music associated with well-known composers like Nadeem-Shravan, Pritam, or A.R. Rahman. The vocal delivery and melodic sensibility didn’t match the genre’s real characteristics at all. A second attempt targeting cumbia, the Latin American dance genre rooted in Colombian tradition and built around accordion, brass, and a driving rhythmic groove, fared no better. Despite prompting for those specific instrumental elements, the output didn’t capture the genre’s actual feel. The result sounded generic rather than rooted in the musical tradition it was asked to emulate. The pattern suggests the model’s training data, or at least its current tuning, skews heavily toward English-language, Western-style pop and adjacent genres. Anyone hoping to use it for regional or non-English music production should expect to do significant work to get usable results, if it’s achievable at all right now. ## Is MiniMax Music 3 worth using right now? For English-language pop and similar genres, it’s a genuinely capable tool: full song structure, stable audio quality across multi-minute tracks, and fine control over arrangement and vocal style through studio mode. The underlying architecture, combining a large structural model with a smaller detail model and a dedicated audio decoder, is a sensible design that shows in the coherence of the output. For anything outside that lane, especially Bollywood, Latin American, or other non-Western genres, the model currently falls short. If your use case involves multilingual or culturally specific music generation, this isn’t yet the tool for that job. It’s a strong English pop generator wrapped in a model that oversells its general-purpose range. ## Frequently Asked Questions ### What audio format does MiniMax Music 3 output? It generates 32 kHz, 16-bit stereo WAV files, with full song structure including intro, verses, chorus, bridge, and outro rather than short loops. ### How much VRAM does MiniMax Music 3 need to run locally? In testing on an RTX 6000 with 48 GB of VRAM, actual usage during generation measured just over 27 GB, suggesting it could run on smaller cards with less headroom. ### Can MiniMax Music 3 generate songs in languages other than English? It’s not officially confirmed as multilingual, and hands-on testing with Bollywood-style and cumbia prompts produced results that didn’t match those genres’ actual musical characteristics. ### What’s the core architecture behind MiniMax Music 3? It uses a hierarchical autoregressive design: an 8-billion-parameter global LLM handles song structure, a 0.6-billion-parameter local LLM fills in acoustic detail, and a flow-matching based variational autoencoder renders the final waveform. ### Does MiniMax Music 3 write its own lyrics? Yes, if you don’t supply lyrics, the model will generate its own based on your genre and mood description, though you can also write full lyrics yourself in studio mode for more control.