thepatch/mdx23c-drumsep-GGUF

DrumSep (MDX23C, aufr33 & jarredou) GGUF

No license is stated for these weights anywhere upstream. aufr33 and jarredou's DrumSep model was released without one; its original release page is no longer online, and the MSST-WebUI mirror it is taken from states none for the weights (its own code is AGPL-3.0). It is republished here in GGUF form with that caveat. Do not assume any license, and if you hold the rights and want it removed, open a discussion on this repository.

aufr33 and jarredou's six-stem drum separation model (MDX23C, ep_141, SDR 10.8059), converted to GGUF for stems.cpp, a C++/ggml stem separator that runs on CPU, CUDA, Vulkan and Metal with no Python. It splits a drum track into kick, snare, toms, hh, ride and crash in one pass.

file	size
mdx23c_drumsep-0.1B-v1.0-F32.gguf	417.3 MiB

Download and run

git clone --recursive https://github.com/betweentwomidnights/stems.cpp && cd stems.cpp
./build.sh vulkan                          # or cpu | cuda | metal   (Windows: build.cmd)
./models.sh mdx23c_drumsep                 # stems.cpp v0.1.4 and later
build-vulkan/bin/stems-split -m models/mdx23c_drumsep-0.1B-v1.0-F32.gguf -i drums.wav -o kit/

stems-server finds it by model name ("model": "mdx23c_drumsep"). It expects a drum stem, for example the drums output of htdemucs.

What the runtime does

The network is ZFTurbo's TFC_TDF_net (v3): InstanceNorm / exact GELU / convolution blocks with 1x1 shortcuts and bottleneck TDF layers, concatenated skips, 5 scales of 2 blocks, 128 channels and 4 frequency subbands, on an n_fft 2048 / hop 512 STFT cropped to 1024 bins. Chunking follows ZFTurbo's generic demix with the config's 4-way overlap. See stems.cpp's docs/MDX.md.

Parity

SNR of the C++ output against ZFTurbo's model code and demix, unchanged, run in float32 on CPU, on a 20 s clip (demucs' test.mp3), as seg (the first chunk) / full (the whole clip), per stem. Measured 2026-10-09 with ggml 9d0d910b and 4ad3b30b (identical). CPU: Core Ultra 9 275HX. Vulkan and CUDA: RTX 5070 Laptop. Cosine is 1.0000000 on every stem.

backend	seg, worst to best	full, worst to best	20 s clip
CPU	94.7 - 123.9 dB	115.3 - 130.8 dB	72.9 s
CUDA	94.4 - 123.0 dB	116.3 - 130.5 dB	8.3 s
Vulkan	96.2 - 124.8 dB	119.0 - 131.8 dB	8.4 s
Metal (Apple M4)	97.8 - 129.0 dB	123.4 - 131.8 dB	24.2 s

On Apple Metal it needs stems.cpp v0.1.5 or later: v0.1.4 aborts there with unsupported op 'PAD'. The Metal row is against references dumped on the M4 (ggml 4ad3b30b, 2026-10-10), peak memory 762 MB. The M4's CPU runs it at about 0.1x realtime, so expect that on Intel Macs.

Sources

• Upstream checkpoint: aufr33-jarredou_DrumSep_model_mdx23c_ep_141_sdr_10.8059.ckpt from the Sucial/MSST-WebUI mirror at revision 90b617b15bd0dc0b784f3d361faca1b51173fe44, sha256 d2a4aa53eb584d21eead358a4e66d1882ad182911be018f052b5da73be9096d0. Trained by aufr33 and jarredou.
• Configuration: the matching training YAML from MSST-WebUI's GitHub at 39287d46072d2d1d91d9f1518abb1dd9a930e3d7 (configs_backup/multi_stem_models/), sha256 1f019da093523d34b95912add62c5ab5467d5640c3361c0cbf7a3dcab2bdd4ab.
• Converted with stems.cpp's tools/convert_mdx.py --preset drumsep. Tensor names stay PyTorch's and the configuration is stored as metadata. The architecture and weights are otherwise unchanged. Conversion is deterministic: re-running it reproduces this file byte for byte.