MiniMax Open-Sources Music 3 Model for 5-Minute Song Generation

MiniMax officially released and open-sourced the Music 3 music generation model on August 14. Combining an 8B parameter LLM with a DiT architecture, it generates complete 5-minute songs with 32kHz 16-bit stereo output based on lyrics and style descriptions, maintaining long-range structural coherence and vocal characteristics. The model has minimal hardware requirements, running on a single 8GB VRAM card, and quickly topped the Hugging Face trend charts upon release.

Confirmed

• Model Architecture: Combines an 8B parameter global LLM with a 2.7B parameter DiT architecture, built on PyTorch. Some technical details mention the use of a Qwen3 + RVQ autoregressive backbone and a flow-matching DIT/DAV decoder.
• Generation Capabilities: Can generate complete songs up to 5 minutes long based on prompts and lyrics, outputting 32kHz 16-bit stereo audio. The model maintains long-range structural coherence (e.g., intro/verse/chorus/bridge), keeping musical themes, rhythm, vocal features, and arrangement consistent throughout. Users can specify parameters like genre, BPM, key, and emotional dynamics.
• Ecosystem & Hardware Adaptation: Released under the Apache 2.0 license, it has been adapted for 🧨 Diffusers and ComfyUI, and is now available on the fal platform. Hardware requirements are low, requiring only a single 8GB VRAM card, allowing it to be combined with video, image, and LLM tools to build multimodal local workflows.

Timeline

• 08-13: Developers discovered a related PR in the Hugging Face Diffusers repository, and officials had submitted multiple generation samples for audition.
• 08-14: Comfy-Org released the model on Hugging Face, followed by the official announcement of the open-source release and weights. The model quickly shot up the platform's trend charts.

Why it matters

MiniMax Music 3 claims to achieve SOTA-level song generation while remaining friendly to consumer-grade GPUs, providing developers, creators, and the open-source community with a production-ready, all-in-one music model that significantly enriches the local AI tool ecosystem.

2026-08-13 ~ 2026-08-15 · 23 related posts

• Minimax Music 3 Open-Weight Release Imminent as Samples Surface — rerri · 2026-08-13
• [source] MiniMax Music 3 Released with Open Weights: Runs on Consumer GPUs — multimodalart · 2026-08-14
• MiniMax-Music-3 Model Now Available on Hugging Face — LadyQuacklin · 2026-08-14
• MiniMax Releases Music3 Generative Model — Sea_Tomatillo1921 · 2026-08-14
• MiniMax Launches Music 3: Generates Up to 5-Minute Complete Songs — huggingface · 2026-08-14
• MiniMax Launches Music3: Next-Gen Open-Weights Production-Ready Music Model — bennash · 2026-08-14
• MiniMax Releases Music3 Generation Model — Acceptable-Cycle4645 · 2026-08-14
• [source] MiniMax-Music3 Hits Top of Hugging Face Trending for Audio Generation — MiniMaxAI · 2026-08-14
• MiniMax Releases Open-Weights Music Model Music 3, Gets Day 0 Support from SGLang — ying11231 · 2026-08-14
• [source] MiniMax Music 3 Hits ComfyUI: Generates 5-Minute Structured Songs with Lyrics — Comfy-Org · 2026-08-14
• MiniMax-Music-3 Tops Hugging Face Trending Models — Comfy-Org · 2026-08-14
• MiniMax Music 3 Hits fal: Natively Generates 5-Minute Songs — DavidmComfort · 2026-08-14
• MiniMax-Music3 Trends on Hugging Face Spaces — MiniMaxAI · 2026-08-14
• MiniMax Open-Sources Music 3.0: Generate 5-Minute Songs on 8GB VRAM — xiaohu · 2026-08-14
• MiniMax Releases Music Generation Model on Hugging Face; Why No Open-Source? — KyeGomezB · 2026-08-14
• MiniMax Releases Open-Weight Music Model Music 3, Called a Real Alternative to Suno — ccerrato147 · 2026-08-14

7 near-duplicate retellings: multimodalart · PurzBeats · multimodalart · BanghuaZ · NerdyRodent · xiaohu · linoy_tsaban