Despite major advances in computational antibody engineering, no systematic comparison of modern open-source LLM backbone families for antibody sequence generation exists, nor is it known whether architectural differences matter at compact model scales. In this study, five compact transformer variants inspired by prominent open-source LLM families (Llama-4, Gemma-3, DeepSeek-V3, Mistral-3 7B, and Nemotron-3) were trained from scratch on 14.5 million OAS sequences for de novo VH single-domain antibody (sdAb) design. Pre-training yielded uniform generative quality: diversity 0.507–0.516 (CV=0.8%), uniqueness ≈1.0, and novelty 0.925–0.977 (CV=2.2%). Structural assessment confirmed high confidence across all five variants (mean pLDDT 92.43–93.77, 85–100% valid; Kruskal–Wallis H=2.056, p>0.05; N=20), while the baselines performed substantially worse (IgLM 60.46, AbLang 38.48). Fine-tuning across four disease targets (SARS-CoV-2, HIV, HER2, Ebola) produced top candidates that were cross-validated using AlphaFold-2, Boltz-2, RoseTTAFold-2, and ESMFold. Docking yielded ΔG=−36.34 to −65.60 kcal/mol; CDR-H3 novelty against the 14.5M OAS corpus confirmed that 81.1% (811/1,000) of sequences were novel, germline content 77–90%, with no strong MHC-II binders detected. Furthermore, a proof-of-concept Agentic pipeline via MCP with Claude Sonnet 4.6 demonstrates automated profiling of 3D structures, residue analysis, and ATOM fragment analysis across disease targets.