Meta Open Sources AudioCraft Music and Sound Models
The release bundles three audio generation models that Meta says are available for public use under an MIT license.
Updated on Nov. 3, 2026 in Artificial Intelligence
Meta has open sourced AudioCraft, a family of text-to-audio and music models made up of MusicGen, AudioGen and EnCodec. The models, weights and code are available under an MIT license.
Why it matters
Making audio generation tools publicly available lets researchers and developers study and build on them directly. Meta says it hopes open sourcing the models will help the wider community limit or remove dataset bias, since the training data contains a large portion of Western-style music with English text and metadata.
MusicGen is released at 300 million, 1.5 billion and 3.3 billion parameters, and AudioGen at 285 million and 1 billion parameters. The extraction does not include benchmark comparisons against other audio generation systems.
The players
Meta
Meta is the technology company that developed and open sourced the AudioCraft models and the Llama 2 language model.
The details
AudioCraft uses the EnCodec neural audio codec to convert raw audio into discrete tokens, creating a fixed vocabulary of music samples. Autoregressive language models are trained over these tokens to generate new ones, which are then converted back to audio with the EnCodec decoder.
Timeline
Meta open sourced Llama 2 for research and commercial use in mid-July.
Meta reported more than 150,000 Llama 2 download requests in its first week, as of July 28.
The Tech Race
AudioCraft extends the open release approach Meta took with Llama 2 in mid-July from text into audio generation. The move positions Meta's open model strategy against other companies building generative AI tools.
Because the models and code are available under an MIT license, developers can use and modify them in their own projects. Creators and researchers may be able to generate music and sound effects from text prompts with tools built on these models.
The takeaway
Open releases like AudioCraft let outside developers test and improve generative audio systems rather than relying only on the company that built them. Readers following AI tools can watch how the community addresses the dataset bias Meta has acknowledged.
Further reading
For broader context on open AI model releases, explore our Artificial Intelligence coverage.
Source note: This article includes information reported by AI Business.
