Best open-source music and sound-generation models in 2026

Open music generation moved fast in 2026. Several models now write full songs with vocals on a single gaming GPU, one streams music to a MacBook in real time, and two turn a sentence into a sound effect. Quality is a matter of taste and the licenses differ a lot. ACE-Step and HeartMuLa are permissive, YuE2 is non-commercial, MiniMax Music 3 asks for credit and for authorization above US$20 million a year, and Stable Audio 3 caps free commercial use at US$1 million. Read the license line, and ask a lawyer about copyright, because developers describe their training data to very different degrees.

  • List updated November 11, 2026
  1. MiniMax Music 3 Largest

    MiniMax's 11.1B song model: lyrics and a style description in, up to five minutes of 32 kHz stereo with vocals out. It fits a 24 GB GPU in BF16, and the license asks products to show the model name and companies above US$20 million a year to get written authorization.

    • MiniMax
    • 11.1B
    • MiniMax Music3 Community (Conditional)
    • from 24 GB
  2. YuE2 Newest

    M-A-P's 3.6B model that writes an editable score before it renders 48 kHz vocals and accompaniment. On the YuE2 team's own benchmark it scores 6.73 against Suno v5's 6.87 and leads the other open models there, but the weights are CC BY-NC and the card asks for a 24 GB GPU.

    • M-A-P
    • 3.63B
    • CC BY-NC (Non-commercial)
    • from 24 GB
  3. ACE-Step 1.5 Best small model

    ACE Studio and StepFun's MIT-licensed model: full songs in seconds, from about 4 GB of VRAM for the 2B turbo DiT, with covers, repainting and lyrics in 50+ languages. Hugging Face counted about 97,000 downloads of the main repository and 816,000 of the ComfyUI copy in the 30 days to 11 November 2026, and the YuE2 team's benchmark puts its quality below YuE2 and MiniMax Music 3.

    • ACE-Step
    • 4B
    • MIT (Permissive)
    • from 4 GB
  4. HeartMuLa

    HeartMuLa's Apache 2.0 song model with a 3B language model and the 48 kHz HeartCodec: lyrics and style tags in, songs of up to four minutes out at about real-time speed. The YuE2 team's benchmark scores it 6.25, just under MiniMax Music 3, and the separate MuLaCover model is CC BY-NC.

    • HeartMuLa
    • 3.94B
    • Apache 2.0 (Permissive)
    • from 26 GB
  5. Stable Audio 3

    Stability AI's instrumental music and sound-effect family: Medium makes up to 6 min 20 s on about 8 GB of GPU memory, and the Small models run on a laptop CPU. Free commercial use stops at US$1 million in annual revenue, and the repositories are gated.

    • Stability AI
    • 1.4B
    • Stability Community (Conditional)
    • from 4.02 GB
  6. Magenta RealTime 2

    Google's live-music model, which plays rather than writes songs: it streams 48 kHz stereo frame by frame and reacts to text, audio and MIDI in about 200 ms. Small runs on any Apple Silicon MacBook, Base wants an M3 Pro or M2 Max or better, and the weights are CC BY 4.0.

    • Google DeepMind
    • 2.4B
    • CC BY 4.0 (Permissive)
    • from 2.53 GB
  7. MOSS-SoundEffect v2.0

    OpenMOSS's Apache 2.0 1.3B text-to-sound-effect model: ambience, creatures and actions up to 30 seconds at 48 kHz, from English or Chinese prompts. It uses 100 sampling steps and publishes no benchmark or VRAM figure.

    • OpenMOSS
    • 1.3B
    • Apache 2.0 (Permissive)
    • from 13 GB

How we chose

  • Open weights you can download today from the developer's official repository.
  • The license of the weights, not only the code, since music models often add revenue caps, credit lines or non-commercial terms.
  • What the model makes: songs with vocals, instrumental music, live streaming or sound effects are different jobs, so they are described separately.
  • Quality is a matter of taste, so we describe features, lengths and hardware instead of ranking by sound. Benchmark scores are labelled as the developer's own.