LLM-Based Machine Translation
Specialising language models of a few billion parameters for translation through continued pre-training, instruction tuning and reinforcement learning.
Tier Grades
Tier A: Frontier AI labs
feasible
ByteDance's 7B Seed-X is comparable to Gemini 2.5 and GPT-4o, and Tencent's Hunyuan-MT ranked first in 30 of 31 WMT2025 pairs.
Tier B: Indonesian AI labs
feasible
Continued pre-training and RL at 7B, as Seed-X and Tower did, is the scale an Indonesian lab can run.
Tier C: Institution units such as ITB
stretch
A unit can fine-tune a 7B model but the full Seed-X recipe used a 26-author team.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Open 7B translators can be served on premises for translation pilots.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeSeed-X and Hunyuan-MT reach their results at 7B parameters; training compute is not in the abstracts.
- DataSeed-X: monolingual and bilingual data across 28 languages; Hunyuan-MT: 33 languages, first in 30 of 31 WMT2025 pairs.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 26 authors on Seed-X; 7 on Hunyuan-MT; 13 on Tower.
- Time horizonMonths.
- Shelf-life signalsGeneral LLMs matching specialised 7B translators at similar cost
- Pipeline roleTier A releases translation LLMs; Tier B adapts them to Indonesian and regional languages; Tier C evaluates; Tier D deploys them on premises.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
TranslateGemmatranslationexemplar
Hunyuan-MT-7Btranslationexemplar
Seed-X PPO 7Btranslationexemplar
TowerInstruct 7B v0.2translationexemplarHy-MT2 1.8Btranslationproduced byTranslateGemma 4Btranslationproduced byTranslateGemma 12Btranslationproduced byTranslateGemma 27Btranslationproduced byHY-MT1.5 1.8Btranslationproduced byHunyuan-MTtranslationproduced byHunyuan-MT-Chimera-7Btranslationproduced byLinks
R&D TopicLLM-Based Machine Translation
Models 11
Hunyuan-MTTencent, translationproduced byTranslateGemmaGoogle, translationexemplarHy-MT2 1.8BTencent, translationproduced byHunyuan-MT-7BTencent, translationexemplarHunyuan-MT-Chimera-7BTencent, translationproduced byHY-MT1.5 1.8BTencent, translationproduced bySeed-X PPO 7BByteDance, translationexemplarTranslateGemma 12BGoogle, translationproduced byTranslateGemma 27BGoogle, translationproduced byTranslateGemma 4BGoogle, translationproduced byTowerInstruct 7B v0.2Unbabel, translationexemplarPapers
- Is ChatGPT A Good Translator? Yes With GPT-4 As The EngineWenxiang Jiao et al., arXiv preprint, 2023 founding
- Tower: An Open Multilingual Large Language Model for Translation-Related TasksDuarte M. Alves et al., arXiv preprint, 2024 state of the art
- X-ALMA: Plug & Play Modules and Adaptive Rejection for Quality Translation at ScaleHaoran Xu et al., ICLR 2025 (per arXiv comment), 2024 state of the art
- Multilingual Machine Translation with Large Language Models: Empirical Results and AnalysisWenhao Zhu et al., NAACL 2024 (per arXiv comment), 2023 state of the art
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language ModelsHaoran Xu et al., ICLR 2024 (per arXiv comment), 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2507.13618 https://arxiv.org/abs/2507.13618
- arxiv.org/abs/2509.05209 https://arxiv.org/abs/2509.05209
- arxiv.org/abs/2402.17733 https://arxiv.org/abs/2402.17733
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "LLM-Based Machine Translation", https://aurion.semai.id/topics/llm-based-machine-translation/, records of October 2026.