Multilingual models are not as common as English ones. Note that this list is evolving and may not be up-to-date.
| Name | Description | Primary languages | Languages with likely a good support |
|---|---|---|---|
| Llama 3.1 | Llama 3.1 405B is likely the best as of 2024-07-24. It is close to GPT-4o performance in English but not quite in other languages yet. | EN, DE, FR, IT, PT, HI, ES, TH | Likely a wide support. Speaks more coherently in CS. |
| Mistral-Nemo-Instruct-2407 | 12B model trained for multilingualism. Seems to generate well in CS if prompted for. | EN, CN, IT, FR, DE, ES, RU | Likely a wide support. Perhaps 85% of all languages have some support. |
| Qwen2 | Qwen2-72B and various other sizes Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B. | EN, ZH | EN, ZH, DE, FR, ES, PT, IT, NL, RU, CS, PL, AR, FA, HE, TR, JA, KO, VI, TH, ID, MS, LO, MY, CEB, KM, TL, HI, BN, UR |
| Gemma 2 | It seems to be capable in Czech (CS), which a small language. I have seen similar reports for some other languages. Probably all more frequent languages than CS are supported. Based on the Whisper dataset these could be following | EN | ZH, DE, ES, RU, FR, PT, KO, JA, TR, PL, IT, SV, NL, CA, FI, ID, AR, UK, VI, HE, EL, DA, MS, HU, RO, NO, TH, CS. |
| Llama-3 | Llama-3 is a multilingual model trained on over 100 languages. It is not intended for use with other languages than English. | EN | EN, FR, DE, ES, KO, AF, IT, CA, RO, CY, TJ, DA, VI. |
| Nemotron 4 | Nemotron-4-15b, or Nemotron-4-340B-Instruct. I haven’t tested this model. | EN | ES, ID, NL, PL, FR, RU, IT, DE, VI, JA, TR, EL, RO, CS, AR, SV, FA, ZH, HU, KO, UK, NO, FL, HI, DA, BG, HR, SK, TH |
| EagleX 7B 2.25T | Faster inference, but generally a bit lower quality. Not available on OpenRouter. I haven’t tested this model much. | EN |
