🔥 BREAKING: MiniMax just dropped Music 3, its open-weight song machine, and ComfyUI rolled out the red carpet with day-zero support. Yes, you read that right. Your GPU is now officially a recording studio with a questionable work ethic.
ComfyUI has spent the last year as the node-based kitchen sink of the AI creative world. Images? Sure. Video? Obviously. Audio? Well, it got there eventually. Now Music 3 has arrived, and the timeline is officially chaos mode: enabled. Full songs, up to five minutes, generated locally from nothing but lyrics and a description of the vibe you want.
Meet MiniMax Music 3: The Open-Weight Song Machine
MiniMax Music 3 is a state-of-the-art open-weights music generation model, and it is built for complete songs rather than loops or clips. Give it a caption and some lyric lines, and it returns a structurally coherent track in 32 kHz, 16-bit stereo. It holds themes, rhythm, vocal identity, and arrangement progression together across the entire song. No more 15-second jingles that loop awkwardly into eternity.
Want to be the boss of your AI band? The lyrics input accepts explicit section tags like [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo], and [Outro]. You decide the song structure, the model does the session musician thing.
And the control surface goes even deeper with Structured Captions: global metadata (genre, BPM, key, scale, emotional progression, production profile), vocal details (gender, timbre, performance style, harmony, backing vocals, effects), and arrangement details (instruments, section-level evolution, groove, bass, percussion, textures, spatial effects). It is basically a mixing console that speaks plain English.
- Full songs with long-range coherence: native support for up to five minutes without losing the thread on melody, rhythm, or vocal identity.
- Lyrics with section tags: structure your song like a pro, from intro to outro, and the model follows the map.
- Expressive vocals: timbre, harmony, and effects are all steerable, with articulation that stays clean across long passages.
The Nerdy Bits Under the Hood
The architecture is a hybrid of two language models and a continuous synthesis stage. An 8B Global LLM (initialized from Qwen3-8B) models the song's long-range semantic and structural progression, while a 0.6B Local LLM restores fine-grained acoustic detail within each frame. Instead of decoding audio from discrete tokens alone, a synthesis module fuses the hidden states of both LLMs through a 2.4B Flow Matching stage and a 123M Flow-VAE decoder. The result: vocal articulation, instrumental texture, and temporal continuity that token-only systems just cannot match.
ComfyUI v0.33.1 Rolls Out the Red Carpet
ComfyUI v0.33.1 landed on August 13 with native text-to-music support for Music 3, alongside Anima tunes checkpoints, MiniMax H3 Context IR and Regenerate nodes, and new Bria GenFill, Eraser, Expand, and Upscale nodes. The update also ships two brand-new nodes purpose-built for the music workflow:
- MiniMax Music3 Text Encode: caption and lyrics text encoding for MiniMax Music 3.
- Empty MiniMax Music3 Latent Audio: creates an empty audio latent for the exact duration you request.
Under the hood, v0.33.1 adds native CUDA graph support interoperable with dynamic VRAM, keeps dynamic VRAM enabled on WSL, fixes MiniMax Music on non-dynamic VRAM setups, and adds early detection of fused QKV weights. In other words: they fixed the stuff that used to make your PC sound like a jet engine taking off.
Getting started is refreshingly painless. Update ComfyUI to v0.33.1, grab the workflow from the template library, add your lyrics and a music description, and hit run. The weights live on Hugging Face under MiniMaxAI/MiniMax-Music3, and MiniMax even published an official music caption rewriter skill that turns a brief description into a full structured caption with one command: npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter.
The tea, as the kids say: open-weight music generation just took a giant leap ahead of the hosted-only crowd. No subscriptions, no credit meters, no "you have reached your song limit for today." Just you, your GPU, and a [Solo] tag that is about to absolutely rip.
Now if you'll excuse me, I have a node graph to rearrange and a masterpiece to accidentally turn into a synthwave apology anthem. Node wrangling, one checkpoint at a time. 🎸
Comments