Measured language support of 29 LLMs in 100 languages

We measure language support by critical errors and hallucinations. We count a language as supported by an LLM when the translation of our four texts, ~3,000 words, has at most one critical error, no hallucinated segment and at most 60 errors per 1,000 words, weighted by severity. A hallucinated segment carries text that is not in the source: an addition, a corrupted output or an artefact that Intento LQA rates critical.

We test every LLM in all 100 languages. A language where we get no translation from the LLM at all, only text in the wrong language or script, garbage, a text cut off part-way or no output, counts as not supported. A language where we get a translation but have no LQA result for it counts as not assessed.

We set each cut where the results divide. Of the 681 one-step LLM solutions we assessed, 609 have no critical error and 50 have one, and the other 22 have two or more. 657 have no hallucinated segment. Beyond 60 weighted errors per 1,000 words, no three-step solution (translation + LQA + repair) meets general translation quality.

Language support and general translation quality use different thresholds. First we check whether we get a translation at all, then whether the translation meets the three criteria, and last whether it meets general translation quality, with no critical error, at most three major errors and at most 20 weighted errors per 1,000 words. Every translation that meets general translation quality also counts as supported.

The best LLMs support nearly every language where we assess them. Translations from Google Gemini 3.1 Pro Preview meet the criteria in every language but Burmese, from OpenAI GPT-5.5 in all but Burmese, Pashto and Yiddish, and from Anthropic Claude Opus 4.8 in all but Frisian, Sundanese and Tatar. Burmese, Sundanese and Yiddish have the fewest supporting LLMs, two each.

Scope of the language-support assessment

• We assess an LLM in a language only where it is on our shortlist by COMET-Kiwi or MetricX. A language marked not assessed is a language we did not assess, not a language the LLM does not support.
• The criteria count errors in four enterprise texts of ~3,000 words. Other content can show other errors.
• We count hallucinated segments in the segment-level findings of Intento LQA, after removing the findings caused by a translation aligned to the wrong source sentence.
• Segment-level findings are not available for Anthropic Claude Opus 4.8 in Bengali, so we assess it on critical errors and weighted errors alone.
• We follow our classification of models: Translated Lara 2, Widn AI Tower, DeepL Next Gen and Google Translation LLM are LLM-based MT engines, so they are not in this analysis of LLMs.