MiniMax Music 2.6

• 🎵 Generates full songs with vocals and instrumentation from prompts
• 🆕 Adds a dedicated instrumental-only mode
• 🔧 Supports rich structure tags placed within the lyrics
• 🎼 New Cover mode reimagines existing tracks while keeping the melody
• 🎚️ MiniMax frames the release around "Cover Reborn" and redefined bass
• 🌐 Served through MiniMax's music-generation API
• 🏢 Built by MiniMax; released April 2026

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax Music 2.6 is the latest entry in MiniMax's song-generation family, producing complete, mixed tracks—vocals and instrumentation together—from a style prompt and lyrics in a single generation. The model accepts structure tags embedded in the lyrics field to shape arrangement, and it offers a dedicated instrumental mode that produces music without vocals, both exposed through MiniMax's music-generation API.

Compared with its predecessor MiniMax Music 2.5, MiniMax positions 2.6 around new creative modes. The company frames the release under the theme "Cover Reborn," alongside what it describes as a redefined bass response. Both versions are delivered through the same music-generation API workflow, where callers supply a style description and lyrics or toggle instrumental output.

The headline new capability is Cover, which extracts a song's melodic skeleton from uploaded audio and lets you change genre, instrumentation, voice timbre, and lyrics while preserving the original melody. Alongside the release, MiniMax open-sourced three agent-oriented "Music Skills" so AI agents can select modes and call the generation API automatically.

Earlier siblings in the line include MiniMax Music 2.0. Beyond music, MiniMax also ships text models such as MiniMax M3 and speech-synthesis models like MiniMax Speech-02 HD, situating Music 2.6 within a broader multimodal model portfolio.