Best Open-Source AI Music Generators 2026: 6 Options Compared

Open-source AI music generation is young, and licensing is where most people get caught. A project can ship its code under Apache-2.0 while the trained weights, which do the actual work, sit under different terms such as non-commercial. The models in this list mostly state permissive terms, but read the model card for the exact weights you download before you release a track. ACE-Step runs locally with modest graphics memory and supports lyric editing and variations. HeartMuLa moved its code and weights to Apache-2.0. DiffRhythm generates full-length songs and needs at least 8 GB of VRAM with chunked decoding. Mustango adds control over chords and tempo from text. YuE v1 covers lyrics-to-song, while the newer YuE2 comes with company-use restrictions that keep it off this list. InspireMusic is a toolkit, so check its weight licenses before using it. All of them are research-grade: expect to run Python, a graphics card and some patience, and expect results that vary from song to song. If you only want a finished track quickly, a hosted service will be easier. Choose an open model when you need to run offline, fine-tune on your own material or keep your lyrics private.

ACE-Step

Apache-2.0 music foundation model that also does lyric editing, repainting and variations. Free on your own hardware; the README says the 2025 update lowered peak VRAM to 8 GB.

Price: Free (Free to run on your own hardware; the README says an update on 2025-05-10 lowered peak VRAM to 8GB) (ACE-Step pricing, checked 2026-10-22).

Platforms and license: Self-hosted (Python, NVIDIA GPU or Apple Silicon). Apache-2.0 (open source) (license source).

Project activity: 5k GitHub stars · last commit Feb 15, 2026

Best for: The most flexible local song generator.

HeartMuLa

Code and all related model weights are Apache 2.0 per its README. Lazy loading helps on a single GPU, though no minimum VRAM figure is stated.

Price: Free (Free to run on your own hardware) (HeartMuLa pricing, checked 2026-10-24).

Platforms and license: Self-hosted (Python). Apache-2.0 (code and model weights) (open source) (license source).

Project activity: 3.8k GitHub stars · last commit Oct 8, 2026

Best for: Commercial-friendly licensing.

DiffRhythm

Apache-2.0 code and DiT weights for full-length songs. DiffRhythm-base needs at least 8 GB of VRAM, using the chunked option to fit.

Price: Free (Free to run on your own hardware; the README says DiffRhythm-base needs at least 8 GB of VRAM (use the chunked option)) (DiffRhythm pricing, checked 2026-10-24).

Platforms and license: Self-hosted (Python). Apache-2.0 (code and DiT weights) (open source) (license source).

Project activity: 2.4k GitHub stars · last commit Nov 27, 2025

Best for: Full songs on a mid-range GPU.

Mustango

MIT code with Apache-2.0 weights on Hugging Face, aimed at controllable text-to-music. The last repository push was June 2025, so updates are rare.

Price: Free (Free to run on your own hardware) (Mustango pricing, checked 2026-10-24).

Platforms and license: Self-hosted (Python). MIT (code); Apache-2.0 (weights on Hugging Face) (open source) (license source).

Project activity: 396 GitHub stars · No commits since Jun 2, 2025

Best for: Controlling chords and tempo from text.

Worth knowing: Repository metadata and Hugging Face model card (declare-lab/mustango) checked 2026-10-23. Last repository push was June 2025, so expect little new work.

YuE (v1)

The original YuE v1 weights are Apache-2.0 on Hugging Face and turn lyrics into songs. The newer YuE2 has separate, non-commercial company terms, so use v1 if you need open licensing.

Price: Free (Free to run on your own hardware (v1 release)) (YuE (v1) pricing, checked 2026-10-24).

Platforms and license: Self-hosted (Python). Apache-2.0 (v1 weights on Hugging Face) (open source) (license source).

Project activity: 12.8k GitHub stars · latest release yue2-v0.1.6 (Sep 9, 2026) · last commit Oct 6, 2026

Best for: Lyrics-to-song experiments.

Worth knowing: The current main branch describes YuE2 with separate terms: free for individuals, non-commercial for research, commercial license for companies. The Apache-2.0 claim covers the original YuE v1 weights, per the Hugging Face model card.

InspireMusic

Apache-2.0 code toolkit for music, song and audio generation. Check the weight licenses; the last push was May 2025, so treat it as a research option.

Price: Free (Free to run on your own hardware) (InspireMusic pricing, checked 2026-10-24).

Platforms and license: Self-hosted (Python). Apache-2.0 (code) (open source) (license source).

Project activity: 23 GitHub stars · No commits since May 20, 2025

Best for: Researchers and tinkerers.

Which one to pick

• You want one general local tool: ACE-Step
• Licensing for commercial release matters most: HeartMuLa
• You have an 8 GB graphics card: DiffRhythm
• You want text control over musical structure: Mustango
• You have lyrics and want a sung result: YuE (v1)