We describe a lightweight system for the WMT26 Chinese-Southeast Asian Multilingual Translation Task, covering 14 directions between Chinese and seven Southeast Asian languages. It uses two independently trained Transformer Base models: a one-to-many (OTM) model for translation from Chinese and a many-to-one (MTO) model for translation into Chinese. Each model has 68,727,808 unique parameters. Both use 134,645 real parallel sentence pairs, augmented with 154,452 back-translated pairs for OTM and a 96,000-pair subset with synthetic Chinese targets for MTO. On the 6,859-pair internal test set, the final CTranslate2-based system achieves direction-family macro-average BLEU scores of 43.70 and 49.03; on the FLORES-200 crossdomain test set, they are 17.52 and 26.03, respectively. A CT2-only microbenchmark on a single NVIDIA GeForce RTX 5090 achieves 237.02 and 241.96 sentences per second on presegmented, source-length-stratified development samples, using the median of three steadystate measurements.