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.

This paper presents IIMAS-RAG, our system for SemEval-2026 Task 8 on evaluating multi-turn retrieval-augmented generation. Our approach combines LLM-based query rewriting, hybrid sparse-dense retrieval with SPLADE and Voyage-3-large fused via Reciprocal Rank Fusion, and answerability-conditioned generation with GPT-4.1. The system ranked 4th out of 38 teams in Subtask A (Retrieval) and 13th out of 29 teams in Subtask C (Full RAG). Our results show that query rewriting is the most impactful retrieval component, while generation remains challenging in low-context and partially answerable scenarios.