According to CoinMeta, MiniMax has officially launched the weight of minimax-music3, and simultaneously launched the online demo and comfyui workflows. This model can generate a complete song of up to about 5 minutes in length at a time based on lyrics and music descriptions, producing 32kHz, 16-bit stereo audio. The biggest feature of Music3 is its controllability; users can directly arrange the song structure using tags and specify the genre, BPM, key, emotional changes, etc. Officials claim that the full precision requires less than 24GB of video memory, and it takes about 22GB to unload through CPU; further layered loading is even possible with an 8GB graphics card. ComfyUI already supports this natively, and a int8 model for lower video memory usage is also provided.