Qwen-35B-A3B Sovereign Fusion (RG-CACF)
This model represents a mathematical fusion of three state-of-the-art 35B-A3B Mixture-of-Experts (MoE) finetunes:
- Kwaipilot/KAT-Coder-V2.5-Dev (Weight:
0.4): Agentic SWE-bench and coding specialist. - ornith-ai/Ornith-1.5-35B-A3B (Weight:
0.35): Long-form chain-of-thought, self-improvement rollouts, and deep reasoning. - XLearning-SCU/HERO-35B-A3B (Weight:
0.25): Hierarchical RL token-efficiency and concise execution.
Mathematical Method: RG-CACF
Built via Riemannian Geodesic & Curvature-Adaptive Coherence Fusion:
- Log-Euclidean Manifold Norms: Scales merged along the Riemannian cone of positive definite metrics.
- MoE Router Sharpening: Logit entropy calibrated to prevent uniform expert dispersion.
- Continuous Sigmoidal Coherence Gating: Destructive task interference eliminated without hard masking cliffs.
- Frobenius Energy Conservation: Weight matrices preserved under geodesic norm projection.