<span class="paragraphSection"><div class="boxTitle">Abstract</div>Deep learning has successfully been applied to design <span style="font-style: italic;">cis-</span>regulatory elements (CREs) for a few species, but a broadly applicable platform for generating functional promoters for thousands of prokaryotes remains lacking. In this study, we introduce a language model for prokaryotic CREs, referred to as PromoGen2, to design CREs without prior experimental data. PromoGen2 was pretrained on CREs derived from 17 000 prokaryotic genomes. It achieved the highest zero-shot prediction correlation of promoter strength across species, improving the average Spearman correlation from 0.27 to 0.50 compared to the best baseline, while reducing the number of parameters by 10³. Artificial CREs designed with PromoGen2 demonstrated a 100% success rate in <span style="font-style: italic;">Escherichia coli, Bacillus subtilis, Bacillus licheniformis</span>, and <span style="font-style: italic;">Agrobacterium tumefaciens</span>. Based on PromoGen2, we developed the Promoter-Factory framework to design promoters from unannotated genomes. Experimental validation showed that most of the promoters designed for <span style="font-style: italic;">Jejubacter sp</span>. L23, a newly isolated halophilic bacterium with no available CREs, were active and capable of driving lycopene overproduction. Additionally, we introduced PromoGen2-proka, a taxonomy-aware model for CRE design based on prokaryotic genera. Experimental validation confirmed its reliable success rate. The combined use of PromoGen2-proka and Promoter-Factory offers a broadly applicable tool for designing CREs for prokaryotes, fulfilling the needs of synthetic biology and microbiology research.</span>