Bluesky Thread

Bytedance SEED-X: a 7B that beats Gemini-2.5-pro on language translation

July 18, 2025 View original thread

Bytedance SEED-X: a 7B that beats Gemini-2.5-pro on language translation

- reasoning model
- pre-trained on 6T tokens
- structured like a mistral

it's entire pre-training dataset is oriented around language translation, often using bilingual samples

Translation Performance (BLEURT) Seed-X-7B Gemini-2.5-Pro 79.2 79.279.2794-79.0 78.8 78.5 791 78.6 76.8 DeepSeek-R1 GPT-40 80.3 80.179.8 78.81 78.478.3 77.2 Qwen3-235B-A22B Llama-4-Scout-17B-16E-Instruct Mistral-Large-2411-123B Gemma3-27B Aya-32B LLaMAX3-8B 77.3 77.0 76.6 76.1 78.0 76.677.1 75.475.0 73.2 70 72.8 72.5 72.3 71.1 68.1 75.575.2 73.8 75.5 73.9 74.8 73.4 71.6 70.270.0 70.3 71.7 71.9 70.4 71.1 70.2 67.1 61.7 60 XX2EN EN2XX XX2ZH ZH2XX Figure 1 Benchmark performance on Flores-200 of SEED-X and its counterparts. XX2XX

i have mixed feelings on this

A) it’s just translation, it’s a narrow task

B) it’s translation! you have to deeply understand the topic and have some level of theory of mind in order to do it well

i was an intern for a linguistics org out of college and they were kind enough to force me to do linguistics courses. i did just enough basic translation to appreciate that its not a simple isomorphic mapping between languages

it’s basically art. the literal translation is usually not that good

More like this