Music source separation into drums, bass, other and vocals. Converted to MLX from the maintained Demucs v4 source and the official HTDemucs-ft weights.

Four checkpoints, one per stem, 84 MB each:

Each .json beside a checkpoint is that stem's architecture configuration.

HTDemucs-ft ships a separately fine-tuned checkpoint per stem, and each one emits all four sources — the stem it is named for is the one it was tuned to produce best. Extracting a single stem therefore means downloading one 84 MB file, and a full four-stem separation runs the network four times.

fp32 only; no quantized conversion is published.

MIT, covering both the upstream code and the official HTDemucs-ft weights.