fMAC-v4: Data-Distilled Cognitive Graph
- Architecture: Cellular Nodes with Shared Dense SwiGLU Core
- Distillation Source: Cosmopedia-v2 (Mixtral-8x22B) & SmolTalk (Llama-3.1-70B)
- Training Step: 4000 / 4000
- Loss: 2.4287
- Active Parameters per token: ~34M (out of 56.5M total)
