While everyone is watching language models fight for benchmark positions, Hugging Face's trending list tells a quieter story: MiniMax's Music 3 is still climbing — 19.7k downloads last month, 1,280 likes, and a community that has already produced 22 quantized variants, 18 finetunes, and 34 Spaces apps around it. The open-source bar for complete-song generation is being raised by a release that got almost no headlines.
Two LLMs, one song
Music 3 uses a hierarchical autoregressive design. An 8B Global LLM predicts the first RVQ codebook frame by frame, modeling the song's long-range semantic and structural progression — musical themes, rhythm, vocal identity, and arrangement evolution. A 0.6B Local LLM fills in the remaining acoustic codebooks within each frame, restoring fine-grained sonic detail. The Global LLM is initialized from Qwen3-8B: during training its embedding and output layers are first adapted to semantic music tokens, then the two LLMs are jointly trained. Song structure is controlled through section tags in the lyrics — [Intro], [Verse], [Chorus], [Bridge], [Outro] — combined with a music description covering BPM, key, and instrumentation. The model follows not just a global style but the musical development of the song over time.
The interesting part: synthesis without discrete tokens
Most music generation models stop at discrete RVQ token decoding. Music 3 goes one step further: the synthesis module directly fuses the final hidden states of both LLMs, and these continuous representations carry richer acoustic information — vocal articulation, instrumental texture, temporal continuity. The synthesis path runs hidden-state fusion → Flow Matching (2.4B) → Flow-VAE latent → a 123M-parameter Flow-VAE decoder, outputting 32 kHz, 16-bit stereo WAV. The training tokenizer uses eight RVQ layers: a 16,384-entry semantic codebook first, then seven acoustic codebooks of 1,024 entries each. At inference, waveform synthesis uses only the fused hidden states — the discrete tokenizer decoder is not needed. The Flow-VAE architecture is adapted from MiniMax's speech stack and retrained for the dynamic range and spectral characteristics of music.
VRAM: official two GPUs, community 8GB
Two deployment paths exist. The official SGLang-Omni path needs two CUDA GPUs: GPU 0 runs Qwen3 and eight-codebook autoregressive generation, GPU 1 runs Flow Matching and waveform decoding. The diffusers path is friendlier for individual developers: full precision fits under 24GB of VRAM; with automatic CPU offloading, generation takes about 22GB; streaming the language model layer by layer brings it down to 8GB cards — at the cost of speed. Limits worth noting: inference requires CUDA, only non-streaming generation is supported, text prompts are capped at 5,000 tokens, and audio is capped at 9,000 acoustic frames (25 frames per second).
Where it sits in the ecosystem
The GitHub repo has 753 stars and 61 forks, and ships a music-caption-rewriter skill that expands a brief natural-language description into a three-part structured caption (Global Metadata / Vocal Details / Arrangement) without external APIs. The model card is also candid: section tags and music descriptions provide generative control rather than strict symbolic guarantees — generated tempo, key, and instrumentation may not match every requested detail exactly.
So what — while the open-source language model race compresses into a knife fight, music generation quietly offers a more complete engineering specimen: hierarchical modeling, continuous hidden-state synthesis, and consumer-GPU deployment, each layer a decision you can take apart. For developers building creative tools, this is an open base you can start hacking today (model card: huggingface.co/MiniMaxAI/MiniMax-Music3).