Magenta RealTime 2 โ€” Apple Silicon build (for Serenance)

These are the Magenta RealTime 2 model weights by Google DeepMind, converted and quantized for on-device Apple Silicon inference. They are consumed by Serenance, a Mac-native focus-music app by Arign Ltd, which streams them at first run and generates music entirely on the user's Mac.

Attribution

Model: Magenta RealTime 2, Google DeepMind Original: https://huggingface.co/google/magenta-realtime-2 Model weights ยฉ Google DeepMind / Google LLC, licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/

Modifications by Arign Ltd (2026): the original weights were converted and quantized for on-device Apple Silicon inference (MLX / .mlxfn format, with the MusicCoCa and SpectroStream encoder stacks packaged for local use). No changes were made to the model's behaviour beyond format conversion and quantization for efficiency.

Serenance and Arign Ltd are not affiliated with, or endorsed by, Google.

Notices (retained from the original model card)

These notices are retained from the original Magenta RealTime 2 model card, per CC BY 4.0 ยง3(a)(1):

  • No warranty. The materials are provided "AS IS", without warranties or conditions of any kind, express or implied, to the extent permitted by law.
  • Responsible use. Use the model responsibly; do not generate content that infringes the rights of others.
  • Outputs. Google claims no rights in the outputs you generate with the model.

License

Distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, the same terms as the original Magenta RealTime 2 weights. You are free to share and adapt this material with appropriate attribution.

Contents

  • resources/ โ€” MusicCoCa (text/music encoders, mapper, quantizer, tokenizer) and SpectroStream assets shared by both tiers.
  • models/mrt2_base/ โ€” full-quality tier (Apple Silicon Pro/Max/Ultra).
  • models/mrt2_small/ โ€” compact tier (all Apple Silicon).
Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for arign/serenance-mrt2

Finetuned
(3)
this model