Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Initial public release.

C++ MLX implementation of Magenta RealTime for Apple Silicon (mlx-stream), sustaining >1.0× real-time generation on M3 Pro and faster.
.mlxfn source-graph bundles for the SpectroStream encoder/decoder and the Depthformer temporal/depth steps, loaded via mlx::core::import_function; falls back to a capturing-lambda compile path when bundles are absent.
bf16 Depthformer compute path selectable via --dtype {fp32|fp16|bf16} (codec stays at fp32 for spectral quality).
Cross-chunk encoder pipelining and hot-path scalar / weight-view caching for per-chunk latency wins on the inner depth loop.
Speculative depth decoding driver in generate_tokens for warm chunks.
PortAudio playback path with prebuffering, device selection (--list-devices, --device), and WAV recording (--record).
scripts/download_weights_from_hf.py to fetch the published MLX weight bundle from Hugging Face into .weights-cache/.
scripts/run_reference_stream.py and reference/ venv bootstrap for an isolated upstream JAX/TF baseline (make ref-stream) suitable for A/B audio comparison against mlx-stream.
Catch2 unit tests for dtype handling and pytest coverage for the reference-stream subprocess protocol.
Apache 2.0 licensing with NOTICE attribution covering both the upstream Magenta RealTime code and the redistributed CC-BY 4.0 weight bundle.