We compare translation time per 1,000 source words because the test texts differ in length.
For each model, the fixed time and the time per 1,000 words come from the median time of each of the four texts across its languages, translated without context in the request.
Seven LLMs have a fixed time of 32 to 125 seconds per request: xAI Grok 4.20 (reasoning), Google Gemini 3.1 Pro Preview, DeepSeek V3.2, Alibaba Qwen3-235B-A22B, z.ai GLM-5, Google Gemma 4 and MoonshotAI Kimi K2.5. With DeepSeek V3.2, the time does not grow with the length of the text.
Fixed time per request and time per 1,000 words, by model
| Model | Fixed time per request, s | Time per 1,000 words, s | Time per word, 270-word text against the 999-word text | Languages |
|---|---|---|---|---|
| MT engine | ||||
| Yandex Cloud Translate v2 | 1.65 | 1.1 | 2.61 | 83 |
| Elia Machine Translation System | 2.06 | 1.1 | 2.66 | 5 |
| TartuNLP | 2.49 | 2.6 | 1.86 | 7 |
| ALL-Lab | 15.81 | 10.8 | 2.65 | 6 |
| LLM-based MT engine | ||||
| Translated Lara 2 | 0.92 | 0.3 | 3.21 | 97 |
| Supertext LLM | 10.83 | 1.1 | 3.23 | 26 |
| LLM | ||||
| DeepSeek V3.2 | 37.38 | -1.600 | 4.33 | 100 |
| xAI Grok 4.20 (reasoning) | 32.02 | 8.2 | 3.28 | 100 |
| OpenAI GPT-5.6 Luna | 3.48 | 10.2 | 1.64 | 100 |
| xAI Grok 4.20 (non-reasoning) | 2.27 | 16.8 | 1.31 | 100 |
| Cohere Tiny Aya | 1.14 | 17.3 | 0.88 | 69 |
| OpenAI GPT-4.1 | 5.65 | 17.8 | 1.51 | 100 |
| Google Gemini 3.5 Flash | 23.97 | 26.9 | 2.67 | 100 |
| OpenAI GPT-5.6 Terra | 5.42 | 30.5 | 1.48 | 100 |
| Google Gemini 3.1 Pro Preview | 35.66 | 36 | 2.47 | 100 |
| MiniMax M2.5 | 12.36 | 37.4 | 1.71 | 100 |
| Anthropic Claude Haiku 4.5 | 3.46 | 39.9 | 1.17 | 100 |
| Mistral Medium 3.1 | 5.91 | 40.8 | 1.3 | 100 |
| AI21 Jamba Large 1.7 | 7.89 | 50.2 | 1.32 | 100 |
| Cohere Command A Translate | 7.34 | 51.4 | 1.24 | 100 |
| OpenAI GPT-5.5 | 19.19 | 56.2 | 1.85 | 100 |
| Anthropic Claude Sonnet 5 | 11.28 | 60.6 | 1.44 | 100 |
| Mistral Large 3 | 10.56 | 60.8 | 1.33 | 100 |
| Anthropic Claude Fable 5 | 12.17 | 66.4 | 1.49 | 100 |
| OpenAI GPT-5.6 Sol | 12.29 | 73.5 | 1.41 | 100 |
| Anthropic Claude Sonnet 4.6 | 8.58 | 75 | 1.22 | 100 |
| Anthropic Claude Opus 4.8 | 12.19 | 75.8 | 1.43 | 100 |
| MoonshotAI Kimi K2.5 | 125.31 | 77.8 | 2.78 | 100 |
| z.ai GLM-5 | 73.89 | 107.9 | 2.04 | 100 |
| open-weight LLM | ||||
| OpenAI gpt-oss-120b | 0.47 | 1.8 | 1.54 | 100 |
| Swiss AI Apertus-8B-2509 | 7.73 | 36.6 | 1.3 | 99 |
| Meta Llama 4 Maverick | 7.07 | 39.7 | 1.16 | 11 |
| SpeakLeash Bielik-11B-v3.0-Instruct | 8.46 | 46.3 | 1.4 | 30 |
| Google Gemma 4 | 80.4 | 79.8 | 1.99 | 100 |
| Alibaba Qwen3-235B-A22B | 39.76 | 96.1 | 1.74 | 79 |
Fitted on the median time of each of the four texts (270, 730, 979 and 999 words) across the languages of the model, translated without context in the request. Models timed one text at a time, without the models that have one timing for all four texts.