Google’s Magenta team unveiled RealTime 2, a set of open-source models that generate music on the fly.
The new release bundles three neural architectures that run on a single GPU and can respond to live MIDI input with near‑zero latency. All code and pretrained weights are hosted on GitHub, and the models support common DAWs through a lightweight plugin. The team also added a small‑footprint version for CPU‑only devices.
For developers, the update removes the need for cloud inference, lowering cost and data‑privacy concerns. Musicians can now experiment with AI accompaniment without an internet connection, which may boost adoption in rehearsal studios and on‑stage setups.
The move echoes earlier Magenta projects that prioritized accessibility over sheer model size, reminding us that open tools often advance more slowly than flashy cloud services.