MT engine translations take a median of ~2 seconds per 1,000 source words, compared with ~66 seconds for LLMs.
Seven MT engines and three LLM-based MT engines are under the real-time cut in their median language. OpenAI gpt-oss-120b, an open-weight LLM, is under the cut per text (1.87 seconds) and just above it per 1,000 words, since with it a short text takes more time per word.
Median time per 1,000 source words, by model
| Model | Type | Median s per 1,000 words | Median s per text | Languages | Real-time languages |
|---|---|---|---|---|---|
| Google Cloud Advanced Translation API | MT engine | 0.17 | 0.13 | 99 | 99 |
| Microsoft Translator | MT engine | 0.23 | 0.17 | 90 | 90 |
| ModernMT Enterprise Edition | MT engine | 0.42 | 0.31 | 93 | 92 |
| Pangeanic PangeaMT | MT engine | 0.54 | 0.4 | 22 | 22 |
| Google Translation LLM | LLM-based MT engine | 0.63 | 0.47 | 80 | 79 |
| DeepL Next Gen | LLM-based MT engine | 0.74 | 0.55 | 85 | 85 |
| Tarjama NMT | MT engine | 1.02 | 0.76 | 1 | 1 |
| Amazon Translate | MT engine | 1.3 | 0.97 | 69 | 68 |
| SYSTRAN PNMT API | MT engine | 1.68 | 1.25 | 57 | 46 |
| Translated Lara 2 | LLM-based MT engine | 2 | 1.14 | 97 | 85 |
| OpenAI gpt-oss-120b | open-weight LLM | 2.73 | 1.87 | 100 | 63 |
| Widn AI Tower Anthill | LLM-based MT engine | 2.89 | 2.16 | 20 | 7 |
| Yandex Cloud Translate v2 | MT engine | 4.11 | 2.5 | 83 | 13 |
| Oracle Cloud Language Service | MT engine | 4.27 | 3.18 | 21 | 0 |
| Elia Machine Translation System | MT engine | 4.72 | 2.86 | 5 | 0 |
| TartuNLP | MT engine | 6.75 | 4.66 | 7 | 0 |
| Tilde TildeOpen LLM engine | LLM-based MT engine | 7.53 | 5.61 | 28 | 0 |
| Widn AI Tower Vesuvius | LLM-based MT engine | 8.84 | 6.58 | 20 | 0 |
| Baidu Translate API | MT engine | 12.15 | 9.04 | 12 | 0 |
| OpenAI GPT-5.6 Luna | LLM | 15.97 | 11.04 | 100 | 0 |
| Cohere Tiny Aya | LLM | 18.45 | 13.67 | 69 | 0 |
| xAI Grok 4.20 (non-reasoning) | LLM | 20.35 | 14.75 | 100 | 0 |
| Supertext LLM | LLM-based MT engine | 20.41 | 11.62 | 26 | 2 |
| Alibaba Machine Translation API | MT engine | 26.27 | 19.55 | 78 | 0 |
| OpenAI GPT-4.1 | LLM | 28.09 | 19.77 | 100 | 0 |
| ALL-Lab | MT engine | 38.15 | 23.84 | 6 | 0 |
| OpenAI GPT-5.6 Terra | LLM | 39.49 | 28.27 | 100 | 0 |
| Anthropic Claude Haiku 4.5 | LLM | 45.33 | 33.12 | 100 | 0 |
| iFlyTek MT | MT engine | 50.8 | 37.82 | 6 | 0 |
| Swiss AI Apertus-8B-2509 | open-weight LLM | 50.82 | 35.04 | 100 | 0 |
| Meta Llama 4 Maverick | open-weight LLM | 51.07 | 36.45 | 11 | 0 |
| Mistral Medium 3.1 | LLM | 53.48 | 39.02 | 100 | 0 |
| Tencent Cloud TMT | MT engine | 55.2 | 41.1 | 13 | 0 |
| Youdao Translation Service API | MT engine | 56.99 | 42.43 | 92 | 0 |
| NiuTrans | MT engine | 58.17 | 43.31 | 22 | 0 |
| MiniMax M2.5 | LLM | 58.39 | 40.98 | 100 | 0 |
| SpeakLeash Bielik-11B-v3.0-Instruct | open-weight LLM | 58.39 | 42 | 30 | 0 |
| Cohere Command A Translate | LLM | 63.59 | 45.87 | 100 | 0 |
| AI21 Jamba Large 1.7 | LLM | 64.03 | 46.45 | 100 | 0 |
| xAI Grok 4.20 (reasoning) | LLM | 65.02 | 38.63 | 100 | 0 |
| DeepSeek V3.2 | LLM | 65.27 | 40.28 | 100 | 0 |
| Naver Papago NMT Commercial | MT engine | 66.61 | 49.59 | 11 | 0 |
| Google Gemini 3.5 Flash | LLM | 67.63 | 44.11 | 100 | 0 |
| Mistral Large 3 | LLM | 78.68 | 55.83 | 100 | 0 |
| Anthropic Claude Sonnet 5 | LLM | 78.77 | 56.1 | 100 | 0 |
| Anthropic Claude Fable 5 | LLM | 87.56 | 61.84 | 100 | 0 |
| Anthropic Claude Sonnet 4.6 | LLM | 89.72 | 64.86 | 100 | 0 |
| OpenAI GPT-5.5 | LLM | 90.39 | 61.52 | 100 | 0 |
| AppTek NMT | MT engine | 90.6 | 67.45 | 54 | 0 |
| Anthropic Claude Opus 4.8 | LLM | 96.27 | 68.97 | 100 | 0 |
| Google Gemini 3.1 Pro Preview | LLM | 99.24 | 61.66 | 100 | 0 |
| OpenAI GPT-5.6 Sol | LLM | 99.38 | 71.69 | 100 | 0 |
| Alibaba Qwen3-235B-A22B | open-weight LLM | 165.31 | 111.77 | 79 | 0 |
| Google Gemma 4 | open-weight LLM | 224.41 | 145.69 | 100 | 0 |
| z.ai GLM-5 | LLM | 237.07 | 159.67 | 100 | 0 |
| MoonshotAI Kimi K2.5 | LLM | 323.15 | 197.33 | 100 | 0 |
Units — Median s per 1,000 words: s per 1,000 words; Median s per text: s per document
Median over the languages of the model, translated without context in the request. Real-time languages: 2 seconds or less per text (~744 words on average). Models dispatched through the Intento platform are timed from its server-side logs.