InspireMusic

Long-form music generation at 48 kHz is the headline capability of InspireMusic, the music, song, and audio generation toolkit in Alibaba's FunAudioLLM family alongside CosyVoice and SenseVoice. Its pipeline chains three components: audio tokenizers that compress waveforms into discrete tokens, an autoregressive transformer built on a Qwen2.5 backbone that predicts tokens conditioned on text or audio prompts, and a super-resolution flow-matching model that upscales the result into 48 kHz stereo output. Released checkpoints include InspireMusic-Base and InspireMusic-1.5B in 48 kHz and 24 kHz variants for 30-second clips, plus InspireMusic-1.5B-Long for long-form generation up to several minutes. Training and inference scripts ship in the repository, installation is from source in a conda environment or via Docker, and Python 3.8 or newer with PyTorch 2.0.1, CUDA 11.8, and flash-attn 2.6.2 or newer is required, so an NVIDIA GPU is effectively mandatory. The repository code carries the Apache-2.0 license. Open-sourced in early 2025 with about 1,400 GitHub stars, it gives researchers a rare fully open text-to-music stack.