Neural DNA (NDNA) uses a compact learned genome to grow network topology through type-based compatibility rules. In prior work (Sudarshan, 2026), we showed that genomes of 226 to 374 parameters could control connectivity in networks with up to 2.2 million connections. Here we test whether this approach scales to real language models. We apply NDNA to GPT-2 Small (124M parameters), where a genome of 354 parameters controls 35.4 million connections in the attention output projections and feed-forward first layers across all 12 transformer layers, a compression ratio of 99,970:1. The genome and model weights are co-trained from scratch on OpenWebText. The genome discovers a striking stratification: layers 5 through 12 converge to 100% connectivity while layers 1 through 4 are progressively pruned, with layer 4 retaining only 7.7% of connections. Despite permanently disabling one-third of all masked connections, the genome-wired model beats GPT-2's published numbers on WikiText-103 perplexity (36.0 vs 37.5), Penn Treebank perplexity (59.4 vs 65.9), and LAMBADA perplexity (22.2 vs 35.1), while reaching 92% of GPT-2 on HellaSwag and 94% on Children's Book Test. Training reveals interpretable dynamics: the genome over-activates early (83% density at iteration 200), prunes aggressively (layers 1 through 4 go to 0% by iteration 800), then partially resurrects layer 1 (from 0% to 98% hard density by iteration 26,000). These results demonstrate that NDNA's developmental framework scales from toy tasks to production-scale language models, achieving 12x higher compression than any prior genome experiment while producing competitive benchmark performance.

Second paper in the Neural DNA series. Paper 1 established the mechanism across four architectures at small scale. This paper scales NDNA to GPT-2 Small (124M parameters).