Internal Specialization Scales with Model Size: Quantitative Evidence from qwen2.5 3B vs 7B

© Copyright 2026 Chen, Ho Yiing (ORCID: 0009-0006-6816-9891), Independent researcher, charenix.com. All rights reserved by the author. This work is licensed under the Creative Commons Attribution 4.0.

Folk wisdom in the LLM community holds that bigger transformers 'specialize more.' I give that claim a cell-level test. Running identical ten-prompt observation passes through qwen2.5:3B and qwen2.5:7B on a Mercury sensor grid produces directly comparable signature counts at each size.

The 7B model has 2.08x as many cross-lingual physics-topic cells, 1.66x as many Chinese-only sensors, and 1.56x as many English-only sensors. It also has 0.41x as many universal-backbone cells; the smaller side dominates that single category. So topical and linguistic circuits expand with parameter count while the always-on backbone shrinks.

This is a cell-level reading of a claim usually argued through perplexity curves and benchmark scores. Both transformers show the cross-lingual physics-topic group, so the phenomenon is not a 7B-only feature that switches on between 3B and 7B; it is present at 3B and grows roughly with parameter count.

Five of six pre-declared lane categories grow with model size; one shrinks. The shrinking category is the universal backbone.

Companion records: Mercury method (10.5281/zenodo.20313154), Mercury-Viewer dataset (10.5281/zenodo.20313150), Mercury discovery (10.5281/zenodo.20313748).