We describe our submission to the WMT26 Multilingual Instruction Shared Task, which evaluates sub-10B-parameter models on multilingual question answering, open-ended generation, and cross-lingual summarization. We compare four open-weight models and vary the system prompt, reasoning mode, and decoding parameters. On our development set, the tested prompt and reasoning changes have a larger effect on LLM-judge scores than the difference between the two finalist models in their best settings. Gemma 4 E4B obtains the highest overall judge score, whereas Qwen 3.5 9B follows hard constraints more reliably on our internal benchmark. We therefore submit Qwen as our primary system and Gemma as a secondary system. We also introduce a 1,283-example diagnostic benchmark for grounded QA, constrained generation, and abstract writing from ACL papers in 24 languages. Using it for QLoRA fine-tuning does not work in our setup, it lowers validation loss but does not improve quality under our LLM judge. In the official automatic evaluation, our Gemma system ranks first overall and our Qwen system second; Qwen wins QA-context, while Gemma wins QA-OEG and places second in summarization.

This paper describes the CUNI-ÚFAL submission to the WMT26 Terminology Translation Task. We present a model-agnostic pipeline that selects document-relevant terminology for a general-purpose LLM, instantiated with GPT-5.6-Sol for both tracks. For Track1 (explicit dictionary) we compare glossary placement, inline annotation, dictionary filtering, and terminology-aware revision: filtering the dictionary per document improves judged coverage and, slightly, translation quality. Inline term insertion buys a small coverage gain at the cost of fluency. For Track2 (in-domain seed bitexts) we compare presenting them as translation history, extracting a dictionary and reusing the Track1 pipeline, and their hybrid. Lacking development references, we evaluated with surface- and lemma-based term matching plus an LLM judge. Among submitted systems in the official evaluation, our Track1 submission ties for the highest COMET-22 score and obtains the best XCOMET-XXL and MetricX-24 XL scores, and our primary Track2 submission achieves the highest term success.