AI Search tests MiniMax Music 3 with pop rock, punk, R&B, progressive house, jazz, Chinese-language, metal and country prompts. The generated songs maintain recognizable genre structure, vocals and instrumentation across a broad range of styles, including vocal accents and multilingual performances.
The local setup runs through ComfyUI with separate diffusion-model, text-encoder and VAE downloads. Model variants range from a 9.8 GB full checkpoint to a 2.5 GB quantized version, while the smallest demonstrated text encoder is 9.2 GB. AI Search reports that a one-minute song takes roughly three to four minutes on a graphics card with 16 GB of memory.
The workflow is currently limited to text-to-music. Prompts can specify overall genre and mood, vocal character, song arrangement, lyrics, structural markers and a duration of up to five minutes. Seed, sampling-step and guidance controls make results repeatable and let users trade generation time against quality and prompt adherence.
AI Search says the output does not yet match the strongest hosted music generators, but local execution provides offline use and direct control. The video also demonstrates companion open tools for editing or covering tracks, building loop-based arrangements, separating songs into instruments and converting parts into editable MIDI.
Watch on YouTube