AI models are trained on large amounts of text, but the distribution across languages is not equal. English dominates the training data. This has direct consequences for the quality of output in other languages. Which models perform best for non-English content production?
When you produce AI content in German, French, Spanish or Dutch, you use the same models as for English. But the quality differs. The amount of training data per language determines how fluent, accurate and contextually aware the model writes. Here is what you need to know.
Large language models learn from text on the internet, books, academic publications and other sources. English is dominant in this: estimates vary, but English typically represents 40 to 60 percent of the training data for many models.
This means a model has learned more patterns, nuances and variations in English than in Dutch. For a language like Swahili or Welsh, that difference is even greater. This translates directly into quality differences in generated text.
Languages with a large online presence perform better in modern models. This applies to: German, French, Spanish, Italian, Portuguese, Japanese, Chinese (Mandarin), Korean and Arabic. These languages are well represented in training data and models produce relatively fluent and accurate text in them.
Dutch falls into the medium-sized category. Quality is good for standard texts, but models struggle with idioms, regional variants and subtle stylistic differences. Technical jargon and sector-specific terms require extra attention in prompts or post-editing.
GPT-4 and GPT-4o (OpenAI): Strong across a broad range of languages, including less common European languages. Good for production tasks in multiple languages.
Claude (Anthropic): Comparable level to GPT-4 for major languages. Reasonably strong in Dutch and other medium-sized European languages.
Gemini (Google): Google has paid extra attention to multilinguality, partly due to the scale of their search engine data. Gemini performs well in many non-English languages.
DeepL Write: Specifically trained for language improvement in several European languages. Strong for post-editing generated text.
Mistral and Llama 3: Open-source models that perform well in major European languages, but fall short qualitatively for smaller language regions.
For some language regions, specialised models exist that outperform the large generic models on specific tasks.
For Arabic: Models like Jais (trained on Arabic data) outperform generic Western models on specific tasks.
For Japanese and Chinese: The large Western models perform reasonably here, but local models from companies like Baidu or NTT may be better for specific applications.
For Scandinavian languages: GPT-4 and Claude generally perform well, partly because those languages are well represented online.
No benchmark list replaces your own test. Create a representative test set: a number of texts in the desired genre and domain. Have them produced by all candidate models. Have them assessed by a native speaker who does not know which model produced what.
This gives an honest picture of actual quality for your specific use case.
Mach8 has experience evaluating and deploying AI models for multilingual content production. We advise based on the specific languages, domains and quality requirements of your project.
There is no universally best model for non-English languages. Performance depends on the language, domain and type of content. Large generic models perform well for common European languages. For specific regions or smaller languages, targeted evaluation is necessary.
Want to know which model best fits your multilingual content needs? Get in touch with Mach8.
We help you go from strategy to implementation. Schedule a no-obligation call.
Schedule a call