UnimeType Releases Open-Source 0.8B and 4B Models for Mixed-Script Transliteration

UnimeType has open-sourced its Transliteration 2.0 models in 0.8B and 4B sizes under the Apache-2.0 license. The models convert phonetic Latin input into native scripts while keeping English words, code identifiers, and URLs in Latin script. They target bilingual writers using Pinyin, Romaji, Hinglish, Arabizi, and Russian transliteration alongside English. The team says existing IMEs force layout toggling and general LLM translators collapse mixed input into one language, so the models were built for sub-second, output-only inference. Both models are fine-tuned on Qwen3.5 and are also available as browser converters.