Massively Multilingual Machine Translation
Dedicated translation models that cover hundreds of languages, many of them low-resource, in one system.
Tier Grades
Tier A: Frontier AI labs
feasible
Meta and Google trained at 200 to 450 languages with teams of 11 to 39.
Tier B: Indonesian AI labs
stretch
Training a 10.7B model on 250 billion tokens is a stretch, though fine-tuning NLLB for Indonesian regional languages is not.
Tier C: Institution units such as ITB
feasible
Small per-pair OPUS-MT models trained on open corpora fit a university unit.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Released NLLB and OPUS-MT English-Indonesian models can be served on premises for document 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
- ComputeMADLAD-400: a 10.7B parameter translation model trained on 250 billion tokens; NLLB uses a sparsely gated mixture of experts.
- DataMADLAD-400: 3T token audited monolingual corpus in 419 languages; NLLB: mined data for 200 languages, evaluated on over 40,000 directions of Flores-200.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 39 listed on NLLB; 11 on MADLAD-400.
- Time horizonYears to build; months to fine-tune.
- Shelf-life signalsLLM translators overtaking dedicated models on low-resource pairs
- Pipeline roleTier A releases many-language models; Tier C and Tier B build Indonesian regional corpora and fine-tune; Tier D serves them for public sector translation.
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
MADLAD-400 MTtranslationexemplar
NLLB-200translationexemplar
OPUS-MTtranslationexemplarMADLAD-400 3B MTtranslationproduced byMADLAD-400 10B MTtranslationproduced by
IndicTrans2 (en-indic 1B)translationproduced byNLLB-200 distilled 600Mtranslationproduced byNLLB-200 distilled 1.3Btranslationproduced byNLLB-200 3.3Btranslationproduced byTransformer (WMT translation, base and big)translationproduced bymBART-50 many-to-manytranslationproduced byOPUS-MT en-idtranslationproduced byOPUS-MT id-entranslationproduced byOPUS-MT en-zhtranslationproduced byLinks
R&D TopicMassively Multilingual Machine Translation
Models 14
NLLB-200Meta, translationexemplarOPUS-MTUniversity of Helsinki (Helsinki-NLP), translationexemplarOPUS-MT en-idUniversity of Helsinki (Helsinki-NLP), translationproduced byOPUS-MT id-enUniversity of Helsinki (Helsinki-NLP), translationproduced byMADLAD-400 3B MTGoogle, translationproduced byMADLAD-400 MTGoogle, translationexemplarNLLB-200 distilled 600MMeta, translationproduced byMADLAD-400 10B MTGoogle, translationproduced byNLLB-200 3.3BMeta, translationproduced byNLLB-200 distilled 1.3BMeta, translationproduced byOPUS-MT en-zhUniversity of Helsinki (Helsinki-NLP), translationproduced byIndicTrans2 (en-indic 1B)AI4Bharat (IIT Madras), translationproduced byPapers
- Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot TranslationMelvin Johnson et al., arXiv preprint, 2016 founding
- Massively Multilingual Neural Machine Translation in the Wild: Findings and ChallengesNaveen Arivazhagan et al., arXiv preprint, 2019 founding
- SeamlessM4T: Massively Multilingual & Multimodal Machine TranslationSeamless Communication et al., arXiv preprint, 2023 state of the art
- MADLAD-400: A Multilingual And Document-Level Large Audited DatasetSneha Kudugunta et al., arXiv preprint, 2023 state of the art
- No Language Left Behind: Scaling Human-Centered Machine TranslationNLLB Team et al., arXiv preprint, 2022 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/2207.04672 https://arxiv.org/abs/2207.04672
- arxiv.org/abs/2309.04662 https://arxiv.org/abs/2309.04662
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Massively Multilingual Machine Translation", https://aurion.semai.id/topics/massively-multilingual-machine-translation/, records of October 2026.