MiniMax Music 3 Open-Sources Weights: SOTA Song Generation on Consumer GPUs

multimodalart · x · 2026-08-14

MiniMax has released its latest music generation model, MiniMax Music 3, with open weights on Hugging Face.

Featuring an architecture of an 8B LLM and a 2.7B DiT, the model can generate full songs from prompts and lyrics, achieving state-of-the-art performance. Notably, it is highly hardware-friendly: via Diffusers or ComfyUI, users can run it locally on small consumer GPUs.

Related event: MiniMax Open-Sources Music 3, Generating 5-Minute Full Songs(13 posts)→

Original post →

More from Multimodal

Multimodal channel →