This paper describes Tilde’s submission to the unconstrained track of the WMT~2026 Model Compression task in the Czech-German direction. Our baseline is TildeOpen-15B-64k, a model distilled from the TildeOpen-30B-64k European foundation model. We adapt it for CS-DE translation, using supervised fine-tuning followed by GRPO-based reinforcement learning. From this baseline, we explore two axes of compression: post-training quantisation and knowledge distillation. For quantisation, we use LLM Compressor to produce a GPTQ-calibrated model with 4-bit NVFP4 weights (10.1\,GB, $\times$3 smaller than the 30.3\,GB footprint of the 15B baseline) and a data-free RTN model with FP8 weights and activations (16.3\,GB). In parallel, we prune and distil the base 15B model into an 8B student and fine-tune it with the same supervised fine-tuning recipe. Applying the two quantisation schemes to the 8B model yields our smallest models at 8.9\,GB and 5.7\,GB. The latter is over $\times$5 more compact and $\times$1.8 faster than the 15B baseline. Evaluation on four test sets across a variety of metrics shows that quantisation and distillation have minimal impact on translation quality. The NVFP4-quantised 15B model offers the best quality/size/speed trade-off and is our primary submission for this competition. All six models are publicly released.

This paper describes Tilde’s submissions to the WMT 2026 Shared Task on Terminology. For both subtasks, we use TildeOpen-30B, which we instruction-tune for translation using terminology glossaries and translation memories, and we detail the construction of the training data augmented with these resources. For Task 1, we compare three term recognition methods aimed at reducing large term collections to subsets relevant for translation: fuzzy search, stemming with exact-match search, and stemming with fuzzy search. For Task 2, we compare extracting a glossary from the provided translation memory and integrating it in context against using retrieved translation-memory entries directly as few-shot examples. As the shared task provides neither reference translations nor a development set, we validate our design decisions with an LLM-as-a-judge protocol. In-domain term collections reduce translation errors by up to 22% relative to translating without terminology, whereas out-of-domain collections yield only marginal gains, and glossaries extracted from the translation memory outperform few-shot translation with retrieved entries.

Large language models often underperform in many European languages due to the dominance of English and a few high-resource languages in training data. This paper presents TildeOpen LLM, a 30-billion-parameter open-weight foundational model trained on 34 European languages to promote linguistic equity and improve performance for low-resource languages. To address the data imbalance, we combine dataset upsampling with a curriculum-based training schedule that alternates between uniform and natural language distributions. The resulting model performs favorably compared to other multilingual LLMs despite being trained with significantly fewer computing resources. Evaluation across multiple multilingual benchmarks shows that TildeOpen surpasses existing open-weight models in text generation and comprehension, particularly for Baltic, Finno-Ugric, and Slavic languages. Human evaluations confirm an up to tenfold reduction in linguistic errors relative to leading baselines. The model and associated resources are fully open-weight and publicly available at huggingface.co/TildeAI/TildeOpen-30b. These outcomes demonstrate that careful data curation and balanced training strategies can substantially enhance multilingual model quality without increasing model size or training volume.