Model artefacts v3: stemd RoFormer

stemd's own separation model. Quality runs it from this release on.

file	sha256	bytes
stemd_roformer.safetensors	ce3802b634d04d955626cffe038ec5ac8f26ef1c90f3d4cfe5ac542ad6da4b37	250,585,752

stemd pins that digest and re-checks it on every load, so a mismatch fails the
install rather than loading whatever the network returned.

Presets. Quality is stemd_roformer alone: one 4-stem BS-RoFormer that
separates drums and vocals in one pass, with harmonics as the remainder. It
replaces bs_polarformer plus htdemucs_ft's drums specialist. Fast and
Balanced still download from models-v2, which keeps every earlier file.

On the 12 MUSDB18-HQ test tracks it was not trained on, through stemd on an
M1 Pro:

	drums	harmonics	vocals	mean	s per track
bs_polarformer + htdemucs_ft	8.79	9.89	12.49	10.39	111.9
stemd_roformer	11.07	11.54	11.34	11.32	60.6

Global SDR in dB. See docs/evaluation.md.

Provenance

Trained for stemd. BS-RoFormer architecture: dim 256, depth 12, 62 bands,
rotary position, four mask estimators (bass, drums, other, vocals), 125.2M
parameters, stored as float16. Converted by
tools/export/convert_stemd_roformer.py and tools/export/convert_roformer.py
into the tensor names of ZFTurbo/Music-Source-Separation-Training's
BSRoformer, into which it loads strictly with the config the converter writes.

Terms

Copyright 2026 Nicolas Saintot.

The weights in stemd_roformer.safetensors are licensed under the
Creative Commons Attribution-NonCommercial 4.0 International
licence (CC BY-NC
4.0). In short:

• You may copy, redistribute, convert and fine-tune the weights.
• Credit stemd, link to this licence, and say if you changed them.
• You may not use the weights, or a model derived from them, for commercial
  purposes. Selling them, bundling them in a paid product, or offering a paid
  service built on them needs written permission.

These terms cover the model only. They place no restriction on the audio
produced with it: stems separated by stemd, or by any other software running
these weights, belong to whoever holds the rights in the source recording, and
may be used however those rights allow.

The licence applies to these weights alone. stemd's code is MIT or Apache-2.0,
at your choice; the other models stemd downloads carry their own terms, listed in
docs/models.md.

Training data includes MUSDB18-HQ (train split), MoisesDB, Slakh2100 and CC0
recordings. Some of these datasets allow non-commercial research use only,
which is why the weights are non-commercial. No MUSDB18-HQ test track was used
for training.

The weights are provided as is, without warranty of any kind, and the licensor
is not liable for any damages arising from their use, as set out in sections 5
and 6 of the licence.

To ask about commercial use, open an issue on
nsaintot/stemd.