Sakhanda Wire
NVDA $152.08 -2.60% MSFT $505.84 -2.35% GOOGL $279.73 -1.41% META $618.09 -1.33% AMZN $256.12 -0.63%
← Back to the news

MiniMax releases open-weights model for five-minute song generation

MiniMax has released MiniMax-Music3, an open-weights text-to-music model that creates complete songs of up to five minutes in one generation. It accepts lyrics containing section tags such as [Verse] and [Chorus], together with a structured caption describing global metadata, vocals and arrangement. The output is a 32 kHz, 16-bit stereo WAV file.

The system combines an 8B Global LLM with a 0.6B Local LLM and a continuous synthesis stack consisting of a 2.4B flow-matching module and a 123M Flow-VAE. Its tokenizer uses eight residual vector-quantization layers. Synthesis uses fused hidden states from the language models and does not load the discrete tokenizer decoder during inference.

MiniMax published the model weights, inference code and three serving options on August 13, 2026. SGLang-Omni uses two CUDA GPUs, while the diffusers pipeline can run below 24 GB of VRAM and as low as 8 GB with group offloading. ComfyUI also provides a native template. Commercial use is allowed under the Community License, with visible “MiniMax-Music3” attribution; products generating more than US$20 million in aggregate annual revenue require prior written authorization.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news