Latency of the evaluated solutions and which of them reach quality in real time.
We call a solution real-time when we get the translation of a text in 2 seconds or less, ~2.7 seconds per 1,000 words for our average text. In their median language, 10 of the 56 models are real-time, translation with 38 takes more than 12 seconds per 1,000 words, and every language has at least one real-time model.
The median LLM time is ~37 times the median MT engine time. With Google Cloud Advanced Translation API, translation takes a median of 0.17 seconds per 1,000 words, and with MoonshotAI Kimi K2.5, 323 seconds. The fastest LLM is OpenAI gpt-oss-120b, an open-weight LLM, at a median of 2.7 seconds per 1,000 words.
To Youdao Translation Service API, NiuTrans and iFlyTek MT, three of the slowest MT engines, we send the texts in 177 requests, one sentence after another, and their time includes all of them.
We assessed real-time one-step solutions in 84 languages, from MT engines such as DeepL Next Gen, Translated Lara 2 and Microsoft Translator and from OpenAI gpt-oss-120b. Only two meet general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words), DeepL Next Gen in German and Romanian. 4 out of 198 meet both tag handling and formatting. A repair of tags and formatting by code does not add time, but it leaves general translation quality as it is.
The one-step solutions that meet every Standard requirement come from LLMs adjusted with context in the request. The fastest is Google Gemini 3.5 Flash in Indonesian at ~12 seconds per 1,000 source words, and across the 38 languages where such a solution exists, the fastest one takes a median of ~65 seconds per 1,000 words. Within a language, the LLM solutions that meet every Standard requirement take ~1.1 times as long as the other LLM solutions at the median, and longer in 27 of the 38 languages.
In each LLM step, we get the whole text again from the LLM, so a post-edit or a repair takes about as long as the translation. We estimate the LQA step at ~85 seconds per 1,000 words, with all its requests in parallel. Two-step solutions (translation + repair by code) take as long as the same solutions in one step, since the repair of tags and formatting by code does not add measurable time.
In the 42 languages where only two-step or three-step solutions meet every Standard requirement, the fastest of them takes a median of ~190 seconds per 1,000 words, against ~60 seconds per 1,000 words where a one-step solution does.
We implement these workflows in Intento Language Hub. Timings from platform logs cover the model calls made through the Hub. The Intento LQA duration here remains an estimate.
We compare the fastest solution that meets the requirements of each quality level in the same languages, those where two solutions meet every Advanced requirement. In German, the fastest solution that meets the Basic requirements is real-time (DeepL Next Gen, under 1 second per 1,000 words), and the fastest one that meets the Standard requirements takes 30 seconds per 1,000 words, a two-step solution (translation + post-edit) on OpenAI GPT-5.6 Luna.
Meeting the Advanced requirements takes 2 to 4.4 times as long in 10 languages, such as Bulgarian (26 to 116 seconds per 1,000 words). We check gender neutrality only on the two best solutions of each language, those with the fewest errors, so the fastest solution that meets the Advanced requirements often runs on a slower model or takes more steps.
Median times differ between LLMs by up to ~320 seconds per 1,000 source words. We compare the translation time of the same LLM in the same language without context and with context in the request, for 29 LLMs in 100 languages, and 62% of the translations take longer with context. With MT engines, the time changes by ~0.2 seconds per 1,000 words with the glossary and formality parameters.
The agent works with one-step, two-step (translation + post-edit or repair by code) and three-step solutions (translation + LQA + repair), on the source sentences while we translate them and on the translated sentences in the later steps.
One-step solutions take a median of 63 seconds per 1,000 words without the agent and 64 seconds with it, two-step solutions 98 and 110 seconds, and three-step solutions 223 and 228 seconds (the LQA step not counted).
Odia is among the ten slowest languages for 23 of the 27 models we timed in 80 or more languages, Lao for 22, and Chinese Simplified among the ten fastest for 21. Sorted by translation time, neighboring languages differ by less than 50% for 20 of the 27 models, and the other seven have one or two outliers at the slow end.
The four source texts range from 270 to 999 words. We model translation time as a fixed time per request plus a time per word. With MT engines, most of it is the fixed time (~1.7 seconds per request for Yandex Cloud Translate v2), and with most LLMs it is the time per word (~0.06 seconds per word for Anthropic Claude Sonnet 5), so with an MT engine a short text takes more time per word.
We record the time of each text while we translate it, and for the models we call through the Intento platform, we take it from the platform logs, so these times are indicative.
Leave feedback