Languages
The languages a model was built for.
Many languages
Multilingual models are trained across dozens or hundreds of languages and can answer, translate and switch between them. Because the web is dominated by English, builders must sample other languages more heavily, collect new data and design tokenizers that do not split non-Latin scripts into too many pieces
English first
English-first models were trained mostly on English text, because English dominates the web, books, code comments and the research papers that set the agenda. Most early language models, from BERT and GPT-2 to Chinchilla and the first Llama, fall in this group, and many small and research models still do
Chinese
Chinese-language models are built for Chinese as well as English, and the strongest now compete with the best models anywhere. China's labs, including Alibaba, DeepSeek, Zhipu (Z.ai), Moonshot AI, Baidu, ByteDance, Tencent, Huawei, MiniMax and iFlytek, released so many models in 2023 that Chinese media called it a battle of a hundred models, and families such as Qwen and DeepSeek are among the most downloaded open models in the world.
Indian languages
Indic models are built for the languages of India: Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Gujarati, Kannada, Odia, Malayalam, Punjabi, Assamese and others, 22 of them listed in the Eighth Schedule of the Constitution. Many Indians also mix languages, as in Hinglish, and type Indian languages in the Latin alphabet
Arabic
Arabic models are built for Arabic, spoken by over 400 million people across more than 20 countries in many dialects, alongside the formal Modern Standard Arabic. Arabic is under-represented online relative to its number of speakers, is written from right to left and has many word forms built from each root, so general models have often handled it poorly
European languages
European-language models are built for the languages of Europe beyond English: French, German, Italian, Spanish, Polish, Portuguese, the Nordic languages and the 24 official languages of the EU in general. Some come from companies such as Mistral AI in Paris; many are public projects trained on Europe's supercomputers, with a stress on openness, data provenance and the EU AI Act.
Japanese
Japanese models are built for Japanese, which mixes three scripts (kanji, hiragana and katakana), leaves no spaces between words and depends heavily on politeness levels and context. Japan has many builders: universities and national research institutes, telecoms and electronics groups, and start-ups such as Sakana AI and Preferred Networks, supported by a government programme that provides computing time.
Korean
Korean models are built for Korean, written in the Hangul alphabet grouped into syllable blocks, with word endings that change with politeness and grammar. South Korea's internet companies and conglomerates, led by Naver, LG, SK Telecom, KT and Kakao, along with start-ups such as Upstage, train their own models, and the government now funds a national programme.
Other languages
This group covers models for the many languages outside the big families: the languages of Southeast Asia, Africa, Latin America, Central Asia and the Pacific, as well as minority and indigenous languages. Most have little text online, so builders gather data from radio, books and volunteers, and often adapt an open model rather than train one from scratch.