Technical Note
Local Execution and Hybrid Topology Checkpoint for Nemotron-3 Super-120B
Local execution and canonical mixed-edge topology for a 120B-parameter hybrid model.
What this proves
NVIDIA-Nemotron-3-Super-120B-A12B-BF16 executes locally under governed streamed execution with a frozen anchor, canonical topology emission, and determinism-gated admission.
The admitted topology surface combines 880 expert-routing edges and 16
attention edges across 88 layers (40 Mamba2, 40 MoE, 8 full-attention),
producing a mixed-edge
topology.v2
artifact with 896 bounded edges.
All topology phases are reproducible across independent runs and pass the determinism gate. One bounded observation is derived from the expert-routing surface.
What this does not prove
A bounded Mamba state-transition surface. Mamba2 layers contribute topology nodes but no edges in the current surface.
Cross-backend, Metal, or browser execution. Results are from a local CPU-only admitted measurement surface.
Cross-platform reproducibility beyond a single Apple M4 Pro.
Semantic or mechanistic conclusions from the expert-routing observation.
Long-context or multi-token generalization beyond the bounded setups.