Meta today released AudioCraft, a framework that generates music and sound effects from text descriptions, and is open-sourcing the code under MIT and the model weights under CC-BY-NC. The toolkit bundles three models: MusicGen, trained on 20,000 hours of Meta-owned or licensed music; AudioGen, trained on public sound effects; and an improved EnCodec decoder for higher-quality music generation with fewer artifacts.

The release is notable for its accessibility — AudioCraft works from raw audio signals rather than MIDI, and the single codebase supports training custom models. Some commentators see the open-source approach as a contrast to closed models like those from OpenAI, and Meta positions AudioCraft as a tool for musicians, game developers, and small businesses.

On Hacker News, the announcement drew 906 points and 319 comments. Commenters debated whether the CC-BY-NC license on the weights truly qualifies as open source, with some arguing that the noncommercial restriction makes it effectively unusable for most real-world applications. Others praised the move as a significant step for audio AI research, noting that the code is MIT-licensed even if the weights are restricted.

The community response captured the week’s tension: Meta wants to be seen as the open alternative in AI, but critics argue that open weights with a noncommercial clause do not match the spirit of open source.