We present the TinyTitans submissions for the WMT26 Model Compression shared task, which focuses on Czech-to-German translation. Starting with the instruction-tuned Gemma 3 12B model, we compare two approaches: 4-bit GPTQ, which uses task-specific calibration examples, and a more aggressive low-rank factorization method that replaces attention and feed-forward projections with truncated Singular Value Decomposition (SVD) factors. On a 300-sentence development set, the calibrated GPTQ retains 99.7% of the dense model’s COMET score while reducing GPU memory usage from 26.6 GiB to 17.6 GiB. The recovered SVD model reduces the number of parameters by 46.6% and reaches 0.7804 COMET in a separate evaluation using 1,997 sentence pairs. In an external evaluation using 457 WMT25 paragraphs, the GPTQ, SVD, and SVD-plus-quantization submissions received COMET scores of 0.5751, 0.4263, and 0.3354, respectively. These findings suggest that while post-training quantization effectively reduces memory usage, combining low-rank truncation with low-bit quantization may increase approximation errors without clear efficiency gains.

We present a new approach called MeritOpt based on the Personalized Federated Learning algorithm MeritFed that can be applied to Natural Language Tasks with heterogeneous data. We evaluate it on the Low-Resource Machine Translation task, using the datasets of South East Asian and Finno-Ugric languages. In addition to its effectiveness, MeritOpt is also highly interpretable, as it can be applied to track the impact of each language used for training. Our analysis reveals that target dataset size affects weight distribution across auxiliary languages, that unrelated languages do not interfere with the training, and auxiliary optimizer parameters have minimal impact. Our approach is easy to apply with a few lines of code, and we provide scripts for reproducing the experiments (https://github.com/VityaVitalich/MeritOpt).