dgrx-systems/Spectra-Mamba-KAN-Hybrid-Language-Model

Spectra: A Mamba-KAN Hybrid Language Model

A complete architectural, mathematical, and implementation reference for Spectra — a hybrid language model that fuses Selective State Space Models (Mamba) with Kolmogorov-Arnold Networks (KANs) in a parallel block design with learnable mixing weights.

Key properties:

• 🚀 O(L) complexity — eliminates the quadratic attention memory wall
• 🧠 Constant inference memory — SSM state replaces the growing KV cache
• 🔍 Interpretable — every KAN edge is a plottable univariate B-spline
• 📐 339M parameters at base configuration (d_model=768, 24 layers)

---

---

Files

• Spectra_Architecture_fixed.pdf — full downloadable document

Citation

@misc{spectra2024,
  title  = {Spectra: A Mamba-KAN Hybrid Language Model},
  year   = {2024},
  note   = {Architecture reference document, MIT License}
}