Pretrained NeuroMamba checkpoints (Aurora-trained, ALCF/NERSC compute) for the NeurIPS 2025 BrainBodyFM workshop paper.
Three model variants (headdim=embed_dim, n_heads=2, d_state=embed_dim/4, expand=2, 12 Mamba2 layers):
| Asset | Parameters | embed_dim | Size |
|---|---|---|---|
1M_E128.pt |
1.4 M | 128 | 25 MB |
3M_E192.pt (paper main) |
3.1 M | 192 | 42 MB |
5M_E256.pt |
5.4 M | 256 | 61 MB |
See README §Pretrained Checkpoints for end-to-end loading and the four fine-tuning scenarios. Paper: https://openreview.net/forum?id=kftg4lmQi8