MiniMax has made available MiniMax‑Music3, an open‑weights system for text‑to‑music generation. It accepts lyric text with section tags and a separate structured caption, and produces a 32 kHz, 16‑bit stereo WAV file up to five minutes long in a single forward pass. Its core is a Hybrid‑LM consisting of an 8 billion‑parameter global language model paired with a 0.6 billion‑parameter local model.
The model uses an eight‑layer RVQ tokenizer with a 16 384‑entry semantic codebook and seven 1 024‑entry acoustic codebooks. Training first optimizes the semantic layer, then jointly tunes all eight. The global model predicts the first RVQ frame per step for long‑range structure; the local model fills the remaining codebooks within each frame. Final hidden states condition a 2.4 B flow‑matching module, whose latent output is decoded by a 123 M Flow‑VAE from MiniMax Speech; at inference the discrete RVQ decoder is not loaded.
- Model: 8 B global + 0.6 B local LLMs, 2.4 B flow‑matching, 123 M Flow‑VAE; RVQ tokenizer (1×16 384 semantic, 7×1 024 acoustic)
- Serving: Diffusers ≤24 GB VRAM (≈22 GB CPU offload, ≈8 GB group offload); SGLang‑Omni split‑GPU; ComfyUI FP16/INT8 template; open‑weights license
Why this matters
Splitting language modeling into global and local components and coupling them to a continuous flow‑matching synthesizer offers a route to improve long‑range musical coherence without relying on discrete token decoding. Released under an open‑weights license with multiple serving options, the system can be experimented with or deployed on a range of hardware.