STEMMA-MF2601

📄 Paper  |  💻 Code  |  🤗 STEMMA-Instruct  |  🤗 STEMMA-Bench

nvidia/music-flamingo-2601-hf fully fine-tuned on STEMMA-Instruct for multi-audio music reasoning, from the paper STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models. Inference uses native Transformers/PyTorch code included in this repository.

Usage

pip install -r requirements.txt
python examples/generate.py --model . --audio a.wav b.wav --prompt "Audio A: <audio>
Audio B: <audio>
Compare A and B."

import sys
from huggingface_hub import snapshot_download

folder = snapshot_download("hoyso48/STEMMA-MF2601")
sys.path.insert(0, folder)
from stemma import Request, load_model

model = load_model(folder)
print(model.generate([Request("Audio A: <audio>\nAudio B: <audio>\nCompare A and B.", ["a.wav", "b.wav"])]))

Up to five audio inputs per request; each <audio> marks one input, in order.

License

Model weights: NVIDIA OneWay Noncommercial License (academic, noncommercial use only; licenses/NVIDIA-OneWay-Noncommercial.txt), inherited from the base model. Code in stemma/ and examples/: Apache-2.0 (LICENSE); stemma/vendor/ is adapted from MOSS-Music (Apache-2.0).

Citation

@misc{sohn2026stemma,
  title         = {{STEMMA}: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models},
  author        = {Sohn, Hoyeol and Kim, Wonil and Kim, Keunhyoung and Kum, Sangeun and Kim, Taehyoung and Moon, Dongjoo and Charoenchob, Theerasak and Weerapang, Teeratep and Lee, Jongpil and Nam, Juhan},
  year          = {2026},
  eprint        = {2610.11884},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2610.11884}
}