Compare translation API features

An overview of the features and limitations of the major machine translation APIs

This page was last updated on April 21st, 2021. We generally compile this comparison only from public or publicly derivable information. You can send updates and suggestions to [email protected].

At ModelFront, we're often asked for advice on the capabilities of the major machine translation systems.

Translation quality is highly subjective, and language-, domain- and application-specific. We don't find any single system to be the best in general across all those dimensions. We recommend running a good evaluation on multiple systems with your actual data.

But as important as raw quality - and often a decisivie factor on effective quality - are the features and limitations of each API.

For example, Google Translate supports translation from German to French, and supports customization. But did you know that Google Translate doesn't yet support customization for pairs like German to French? Only ModernMT does.

That's because most APIs still translate via English. And DeepL now supports basic customization with its glossary feature, but its not yet available in the API or apps or plugins that connect to the API.

Language coverage varies a lot even for basic support. If you want to translate between Cantonese and Mayan, or Portuguese and Brazilian Portuguese, you'll have to use Microsoft Translator or DeepL. But if you need Yakut, Udmurt or Papamiento, there's only Yandex Translate.

Last but not least, data rights are an important consideration, and for a good reason. Can the translation API provider use your training data for other purposes? Who can read the text in your API calls? More and more providers offer no-trace - they never store it, so your data is safe even if they are tortured.

Here's our current list of the most popular self-serve translation APIs with public per-character pricing.

{"content":"*This page was last updated on {% include date.html %}. We generally compile this comparison only from public or publicly derivable information. You can send updates and suggestions to [[email protected]](mailto://[email protected]).*\n\nAt ModelFront, we're often asked for advice on the capabilities of the major machine translation systems.\n\nTranslation quality is highly subjective, and language-, domain- and application-specific. We don't find any single system to be the best in general across all those dimensions. We recommend running [a good evaluation](/what-is-a-good-evaluation/) on multiple systems with your actual data.\n\nBut as important as raw quality - and often a decisivie factor on effective quality - are the **features and limitations** of each API.\n\nFor example, Google Translate supports translation from German to French, and supports customization. But did you know that Google Translate doesn't yet support customization for pairs like German to French? Only ModernMT does.\n\nThat's because most APIs still [translate via English](https://modelfront.com/direct-translation/). And DeepL now supports basic customization with its glossary feature, but its not yet available in the API or apps or plugins that connect to the API.\n\nLanguage coverage varies a lot even for basic support. If you want to translate between Cantonese and Mayan, or Portuguese and Brazilian Portuguese, you'll have to use Microsoft Translator or DeepL. But if you need Yakut, Udmurt or Papamiento, there's only Yandex Translate.\n\nLast but not least, data rights are an important consideration, and for a good reason. Can the translation API provider use your training data for other purposes? Who can read the text in your API calls? More and more providers offer no-trace - they never store it, so your data is safe even if they are tortured.\n\nHere's our current list of the most popular self-serve translation APIs with public per-character pricing.\n\n
\n {{ page | jsonify }}\n
\n\n\n\n
\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
Languages CustomizationContextPricing
for 1M chars
Data rights
Default terms for API calls
Google
Translate API
108Terminology

AutoML,
to and from English only
$20
$80 for custom
\n No-trace
LingvaNex
Translate API
105$5\n No-trace
Yandex
Translate API
93Terminology

Custom models
in private beta
$6
₽447
\n No-trace
Microsoft
Translator Text API
90terminology

Custom Translator,
to and from English only,
limited pairs
$10
$40 for custom
\n No-trace
GDPR-compliant
HIPAA-compliant
Amazon
Translate API
73Terminology

Adaptive customization,
to and from English only
$15
$60 for custom adaptive
\n It’s complicated.
ModernMT
API
47Terminology

Adaptive customization
for all pairs
Document-level$15
\n $15 for custom adaptive
\n $50 for real-time adaptive
\n
\n No-trace
GDPR-compliant
IBM
Watson Language Translator
39Terminology

Custom models
$20
$100 for custom
\n Client-only
Stored temporarily
Baidu
翻译API
28Terminology

Custom models

vertical-specific models
¥49
Custom unlisted
\n It’s complicated.
DeepL
API
24€20\n No-trace
Alibaba
Machine Translation
15
Limited pairs
Commerce-specific model$33\n Client-only
Stored temporarily
\n
\n\nOnly features available via API are considered. For example, Google supports gender variants, and DeepL supports context awareness across lines, but those features are only available in their consumer websites and applications, not via their APIs.\n\nLanguage counts consider the listed languages, no matter how similar, and assume essentially all pairs of those languages are supported unless otherwise noted.\n\nPricing considers the paid rate at moderate volume, not free quota or volume discounts, nor prepaid or subscription discounts. Most of the APIs offer free quota, and charge small additional fees for training custom models and keeping them deployed. There is also continued widespread unpaid use of unofficial APIs based on reverse-engineering consumer translation applications.\n\nSome systems use different architectures and even approaches for different language pairs or content types.\n\nFor more details, please see the documentation of each API.\n\n---\n\n##### Other machine translation APIs\nThere are other self-serve APIs to consider, notably those from **Naver**, **eBay**, **Tencent** and **Youdao**. There are also APIs like the **Intento** and the **Rakuten** RapidAPI that aggregate many translation APIs and more.\n\n##### Other machine translation features\nThere are other features to consider, like locale, formality, data privacy, language identification, HTML and template translation, input correction, transliteration, speech input, latency, batching, offline models and on-prem deployment, that are out of the scope of this comparison for now.\n\nFor example, Google Translate produces French mostly with formal *vous*, Parisian French word choices but Canadian French punctuation. Microsoft Translator now offers a choice between Parisian and Canadian French, but if you want informal French with *tu* out of the box, you'll have to make some compromises - none of the major APIs produces it consistently.\n\n##### Other machine translation providers\nSYSTRAN, Unbabel, Lilt, Tilde, Omniscien, KantanMT, PangeaMT, TextShuttle, Globalese, Tarjama, Iconic, AppTek, SAP and more train custom machine translation, but do not offer a self-serve API with public per-character pricing. SYSTRAN offers an open-source library, OpenNMT.\n\n##### Other machine translation players\nApple, Facebook and Rakuten have their own machine translation deployed in their own consumer platforms, but do not offer an official API. Facebook does offer an open-source library, Fairseq, and pre-trained models.\n\nTo the best our knowledge at this time, many others with machine translation in their platforms, like Twitter and AirBnB, as well as translation providers and CAT tools like Lionbridge and SDL, use the APIs listed above or on-premise deployments of other providers and do not develop their own machine translation from scratch.\n\n\n#### Machine translation libraries\n\nMany major machine translation providers also offer open-source libraries and even pre-trained models.\n
\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
OwnersUsersFrameworkLanguage
FairseqFacebookFacebookPyTorchPython
Marian NMTMicrosoftMicrosoft Translator, Unbabel, Rakuten, eBay, Tilde, Project Bergamot, Opus-MTC++
ModernMT
based on Fairseq
ModernMTModernMTPyTorchJava, Python
OpenNMTSYSTRAN, UbiqusSYSTRAN, Ubiqus, TarjamaTensorflow, PytorchPython
Trax TransformerGoogleGoogle TranslateTensorflowPython
SockeyeAmazonAmazon TranslateMXNetPython
\n
\n \nNote that there are many more users of all these libraries and clients of the APIs that use them. It is not possible to track all closed use of open-source code. For example, ModernMT is used by Translated, AirBnb and the European Parliament.\n\n\n#### Pre-trained machine translation models\n\nOpus-MT offers pretrained Marian NMT models for many language pairs.\n\nFacebook offers pre-trained Fairseq models for a limited number of languages.\n\nThey are also available via Hugging Face.\n\n---\n\n### Updates\n\n##### March 2021 - DeepL\nDeepL launches **Bulgarian**, **Czech**, **Danish**, **Estonian**, **Finnish**, **Greek**, **Hungarian**, **Latvian**, **Lithuanian**, **Romanian**, **Slovak**, **Slovenian**, and **Swedish**, bringing its total number of languages to 24.\n\n##### February 2021 - Microsoft Translator\nMicrosoft Translator launches **Albanian**, **Amharic**, **Armenian**, **Azeri Turkish**, **Khmer**, **Lao**, **Myanmar**, **Nepali**, and **Tigrinya**, bringing its total number of languages to 90. It is the first API to support Tigrinya.\n\n##### January 2021 - Microsoft Translator\nMicrosoft Translator launches **Inuktitut** first, bringing its total number of languages to 79.\n\n##### November 2020 - Amazon Translate\nAmazon Translate launches **adaptive** translation as Active Custom Translation, joining ModernMT, and launches more than a dozen new languages, bringing its total number of languages to 73.\n\n##### October 2020 - Microsoft Translator\nMicrosoft Translator launches support for **Canadian French**.\n\n##### September 2020 - Microsoft Translator\nMicrosoft Translator launches **Assamese** first, bringing its total number of languages to 78. It also updated its system for training custom models.\n\n##### August 2020 - LingvaNex\nLingvaNex launches their translation API with **more than 100 languages**, second only to Google Translate.\n\n##### August 2020 - Microsoft Translator\nMicrosoft Translator launches **Odia (Oriya)**, **Dari**, **Pashto** and **Sorani Kurdish** and **Kurmanji Kurdish**, bringing its total number of languages to 77.\n\n##### June 2020 - Microsoft Translator\nMicrosoft Translator launches **Kazakh**, bringing its total number of languages to 72.\n\n##### April 2020 - Microsoft Translator\nMicrosoft Translator launches **Marathi**, **Gujarati**, **Punjabi**, **Malayalam** and **Kannada**, bringing its total number of languages to 71.\n\n##### April 2020 - DeepL\nDeepL launches **Brazilian Portuguese** in addition to Portuguese, joining Microsoft Translator in offering two varieties as a target language. In contrast, Google's only option is Brazilian Portuguese.\n\n##### March 2020 - ModernMT\nModernMT announces **free machine translation** for April and May, unlimited other than by its hardware constraints.\n\n##### March 2020 - Yandex Translate\nYandex Translate launches **Yakut**, bringing its total number of languages supported to 97. However it's one of 4 languages not yet available in the API.\n\n##### March 2020 - DeepL\nDeepL adds **Chinese** and **Japanese**, its first non-European languages, bringing its total number of languages supported to 11. \n\n##### February 2020 - Google Translate\nGoogle Translate adds **Kinyarwanda**, **Odia (Oriya)**, **Tatar**, **Turkmen** and **Uyghur**, bringing its total number of languages supported to 108.\n\n##### January 2020 - Yandex Translate\nYandex Translate launches **Chuvash**.\n\n##### January 2020 - Microsoft Translator\nMicrosoft Translator launches **Irish Gaelic**.\n\n##### October 2019 - ModernMT\nModernMT launches **human-in-the-loop**, **document-level context** and **context adaptation**.\n\n---\n\n### Thanks\n\nWe're very thankful to all who have contributed to this comparison.\n\n","output":null,"collection":"landscapes","url":"/compare","relative_path":"_landscapes/compare.md","previous":null,"excerpt":"

This page was last updated on\nApril\n21st,\n2021. We generally compile this comparison only from public or publicly derivable information. You can send updates and suggestions to [email protected].

\n","next":{"content":"Is a translation good or bad? Measuring quality and risk are fundamental to successful translation at scale. Both human and machine translation benefit from sentence-level and corpus-level metrics.\n\nThe top **production use cases** are **post-editing effort estimation** and **hybrid translation** - safely auto-approving raw machine translation, typically for medium-scale, high-value content like technical support documentation or [product descriptions](/marketplaces). And of course, there have always been tools and processes for final **validation**.\n\nThe top **offline use cases** are [filtering](/filter) parallel corpora - for example, cleaning translation memories - and evaluation - assessing or comparing translation systems or custom models.\n\nMetrics like BLEU are based on string distance to human reference translations and cannot be used for new incoming translations, nor for the human reference translations themselves.\n\nWhat are the **options** if you want to **build or buy** services, tools or technology for measuring the quality and risk of new translations?\n\n---\n\n### Humans\n\nWhether just an internal human evaluation in a spreadsheet, user-reported quality ratings, an analysis of translator post-editing productivity and effort, or full post-editing, professional human linguists and translators are the gold standard.\n\nThere is significant research on human evaluation methods, and quality frameworks like MQM-DQF and even quality management platforms like **TAUS DQF** and **ContentQuo** for standardizing and managing human evaluations, as well as translators and language service providers offering quality reviews or continuous human labelling.\n\n---\n\n### Features\n\nTranslation tools like **Memsource**, **Smartling** and **GlobalLink** have features for automatically measuring quality bundled in their platforms. Memsource's feature is based on machine learning.\n\n\n### Tools and services\n\n**Xbench**, **Verifika** and **LexiQA** directly apply exhaustive, hand-crafted linguistic rules, \nconfigurations and translation memories to catch common translation errors, especially human translation errors.\n\nThey are integrated into existing tools, and their outputs are predictable and interpretable. LexiQA is unique in its partnerships with web-based translation tools and its API.\n\nModelFront [partners](/access/) like **translate5** and GlobalDoc's **LangXpert** integrate ModelFront technology as smart features in their translation systems and language service offerings.\n\n\n---\n\n### Open-source libraries\n\nIf you have the data and the machine learning team and want to build your own system based on machine learning, there is a growing set of open-source options.\n\nThe most notable quality estimation frameworks are **TransQuest** from Tharindu Ranasinghe, **OpenKiwi** from Unbabel and **DeepQuest** from the research group led by Lucía Specia. TransQuest offers pretrained models. **Zipporah** from Hainan Xu and Philipp Koehn is the best-known library for parallel data filtering.\n\nThe owners of those repositories are also key contributors to and co-organizers of the WMT shared tasks on Quality Estimation and Parallel Corpus Filtering.\n\nMassively multilingual libraries and pretrained models like **LASER** are not specifically for translation or translation quality, but a solid baseline for an\nunsupervised approach to parallel data filtering when combined with other techniques like language identification, regexes and round-trip translation.\n\n\n### Internal and bundled systems\n\nUnbabel, eBay, Microsoft, Amazon, Facebook and others invest in in-house machine translation quality estimation research and development for their own use,\nmainly for the content that flows through their platforms at scale.\n\nThe main goal is to safely auto-approve raw machine translation for as much as possible, whether in efficient hybrid translation workflows for localization or customer service, or just to limit catastrophes on user- and business-generated content that is machine translated by default.\n\nTheir approaches are based on machine learning, and they publish papers and code.\n\nKantanQES can provide a quality score with every KantanMT translation - the first machine translation provider to do so.\n\n\n### Systems accessible as standalone APIs, consoles or on-prem\n\n**ModelFront** is the first and only API for translation risk prediction, and it's also based on machine learning. With a few clicks or a few lines of code, you can access a production-strength system.\n\n[Our approach](/our-approach/) is developed fully in-house, extending ideas from the leading researchers in quality estimation and parallel data filtering, and from our own experience inside the leading machine translation provider.\n\nWe've productionized it and made it accessible and useful to more players - enterprise localization teams, language service providers, platform and tool developers and machine translation researchers.\n\nWe built in security, scalability, and support for 100+ languages and 10K+ language pairs, locales, encodings, formatting, tags and file formats, integrations with the top machine translation API providers, and developed a process for [customization](/custom).\n\nWe provide our technology as an [API and console](https://console.modelfront.com) for convenience, as well as on-prem deployments. You can also [access](/access/) our technology through our partners and other projects that we support.\n\nWe continuously invest in curated parallel datasets and manually-labeled datasets and track emerging risks\ntypes as translation technology, use cases and languages evolve.\n\n---\n\n## Build or buy?\n\nBuilding means investing upfront in research and development of an internal system. Key factors in the decision are:\n\n##### Capabilities\nBuild, deploying and maintaining a deep learning-based risk prediction system requires specialized research and engineering experience, as well as data and hardware.\n\n##### Fragmentation and scale\nThe effort required to build and maintain a system or systems is a function of the number of languages, content types and use cases. If there is a single narrow consistent use case with high volume, then building in-house starts to make more sense.\n\n##### Priorities\nMost organizations with those capabilities and scale have many interesting and high-value problems core to their businesses for a team of natural language processing researchers and machine learning engineers to work on.\n\nThe typical organization that chooses to build a risk prediction system in-house has deep experience building and training machine translation itself from scratch, \nand even those organizations also decide to buy external machine translation and external risk prediction for some purposes.\n\n","output":null,"collection":"landscapes","url":"/options","relative_path":"_landscapes/options.md","previous":{"content":"*This page was last updated on {% include date.html %}. We generally compile this comparison only from public or publicly derivable information. You can send updates and suggestions to [[email protected]](mailto://[email protected]).*\n\nAt ModelFront, we're often asked for advice on the capabilities of the major machine translation systems.\n\nTranslation quality is highly subjective, and language-, domain- and application-specific. We don't find any single system to be the best in general across all those dimensions. We recommend running [a good evaluation](/what-is-a-good-evaluation/) on multiple systems with your actual data.\n\nBut as important as raw quality - and often a decisivie factor on effective quality - are the **features and limitations** of each API.\n\nFor example, Google Translate supports translation from German to French, and supports customization. But did you know that Google Translate doesn't yet support customization for pairs like German to French? Only ModernMT does.\n\nThat's because most APIs still [translate via English](https://modelfront.com/direct-translation/). And DeepL now supports basic customization with its glossary feature, but its not yet available in the API or apps or plugins that connect to the API.\n\nLanguage coverage varies a lot even for basic support. If you want to translate between Cantonese and Mayan, or Portuguese and Brazilian Portuguese, you'll have to use Microsoft Translator or DeepL. But if you need Yakut, Udmurt or Papamiento, there's only Yandex Translate.\n\nLast but not least, data rights are an important consideration, and for a good reason. Can the translation API provider use your training data for other purposes? Who can read the text in your API calls? More and more providers offer no-trace - they never store it, so your data is safe even if they are tortured.\n\nHere's our current list of the most popular self-serve translation APIs with public per-character pricing.\n\n
\n {{ page | jsonify }}\n
\n\n\n\n
\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
Languages CustomizationContextPricing
for 1M chars
Data rights
Default terms for API calls
Google
Translate API
108Terminology

AutoML,
to and from English only
$20
$80 for custom
\n No-trace
LingvaNex
Translate API
105$5\n No-trace
Yandex
Translate API
93Terminology

Custom models
in private beta
$6
₽447
\n No-trace
Microsoft
Translator Text API
90terminology

Custom Translator,
to and from English only,
limited pairs
$10
$40 for custom
\n No-trace
GDPR-compliant
HIPAA-compliant
Amazon
Translate API
73Terminology

Adaptive customization,
to and from English only
$15
$60 for custom adaptive
\n It’s complicated.
ModernMT
API
47Terminology

Adaptive customization
for all pairs
Document-level$15
\n $15 for custom adaptive
\n $50 for real-time adaptive
\n
\n No-trace
GDPR-compliant
IBM
Watson Language Translator
39Terminology

Custom models
$20
$100 for custom
\n Client-only
Stored temporarily
Baidu
翻译API
28Terminology

Custom models

vertical-specific models
¥49
Custom unlisted
\n It’s complicated.
DeepL
API
24€20\n No-trace
Alibaba
Machine Translation
15
Limited pairs
Commerce-specific model$33\n Client-only
Stored temporarily
\n
\n\nOnly features available via API are considered. For example, Google supports gender variants, and DeepL supports context awareness across lines, but those features are only available in their consumer websites and applications, not via their APIs.\n\nLanguage counts consider the listed languages, no matter how similar, and assume essentially all pairs of those languages are supported unless otherwise noted.\n\nPricing considers the paid rate at moderate volume, not free quota or volume discounts, nor prepaid or subscription discounts. Most of the APIs offer free quota, and charge small additional fees for training custom models and keeping them deployed. There is also continued widespread unpaid use of unofficial APIs based on reverse-engineering consumer translation applications.\n\nSome systems use different architectures and even approaches for different language pairs or content types.\n\nFor more details, please see the documentation of each API.\n\n---\n\n##### Other machine translation APIs\nThere are other self-serve APIs to consider, notably those from **Naver**, **eBay**, **Tencent** and **Youdao**. There are also APIs like the **Intento** and the **Rakuten** RapidAPI that aggregate many translation APIs and more.\n\n##### Other machine translation features\nThere are other features to consider, like locale, formality, data privacy, language identification, HTML and template translation, input correction, transliteration, speech input, latency, batching, offline models and on-prem deployment, that are out of the scope of this comparison for now.\n\nFor example, Google Translate produces French mostly with formal *vous*, Parisian French word choices but Canadian French punctuation. Microsoft Translator now offers a choice between Parisian and Canadian French, but if you want informal French with *tu* out of the box, you'll have to make some compromises - none of the major APIs produces it consistently.\n\n##### Other machine translation providers\nSYSTRAN, Unbabel, Lilt, Tilde, Omniscien, KantanMT, PangeaMT, TextShuttle, Globalese, Tarjama, Iconic, AppTek, SAP and more train custom machine translation, but do not offer a self-serve API with public per-character pricing. SYSTRAN offers an open-source library, OpenNMT.\n\n##### Other machine translation players\nApple, Facebook and Rakuten have their own machine translation deployed in their own consumer platforms, but do not offer an official API. Facebook does offer an open-source library, Fairseq, and pre-trained models.\n\nTo the best our knowledge at this time, many others with machine translation in their platforms, like Twitter and AirBnB, as well as translation providers and CAT tools like Lionbridge and SDL, use the APIs listed above or on-premise deployments of other providers and do not develop their own machine translation from scratch.\n\n\n#### Machine translation libraries\n\nMany major machine translation providers also offer open-source libraries and even pre-trained models.\n
\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
OwnersUsersFrameworkLanguage
FairseqFacebookFacebookPyTorchPython
Marian NMTMicrosoftMicrosoft Translator, Unbabel, Rakuten, eBay, Tilde, Project Bergamot, Opus-MTC++
ModernMT
based on Fairseq
ModernMTModernMTPyTorchJava, Python
OpenNMTSYSTRAN, UbiqusSYSTRAN, Ubiqus, TarjamaTensorflow, PytorchPython
Trax TransformerGoogleGoogle TranslateTensorflowPython
SockeyeAmazonAmazon TranslateMXNetPython
\n
\n \nNote that there are many more users of all these libraries and clients of the APIs that use them. It is not possible to track all closed use of open-source code. For example, ModernMT is used by Translated, AirBnb and the European Parliament.\n\n\n#### Pre-trained machine translation models\n\nOpus-MT offers pretrained Marian NMT models for many language pairs.\n\nFacebook offers pre-trained Fairseq models for a limited number of languages.\n\nThey are also available via Hugging Face.\n\n---\n\n### Updates\n\n##### March 2021 - DeepL\nDeepL launches **Bulgarian**, **Czech**, **Danish**, **Estonian**, **Finnish**, **Greek**, **Hungarian**, **Latvian**, **Lithuanian**, **Romanian**, **Slovak**, **Slovenian**, and **Swedish**, bringing its total number of languages to 24.\n\n##### February 2021 - Microsoft Translator\nMicrosoft Translator launches **Albanian**, **Amharic**, **Armenian**, **Azeri Turkish**, **Khmer**, **Lao**, **Myanmar**, **Nepali**, and **Tigrinya**, bringing its total number of languages to 90. It is the first API to support Tigrinya.\n\n##### January 2021 - Microsoft Translator\nMicrosoft Translator launches **Inuktitut** first, bringing its total number of languages to 79.\n\n##### November 2020 - Amazon Translate\nAmazon Translate launches **adaptive** translation as Active Custom Translation, joining ModernMT, and launches more than a dozen new languages, bringing its total number of languages to 73.\n\n##### October 2020 - Microsoft Translator\nMicrosoft Translator launches support for **Canadian French**.\n\n##### September 2020 - Microsoft Translator\nMicrosoft Translator launches **Assamese** first, bringing its total number of languages to 78. It also updated its system for training custom models.\n\n##### August 2020 - LingvaNex\nLingvaNex launches their translation API with **more than 100 languages**, second only to Google Translate.\n\n##### August 2020 - Microsoft Translator\nMicrosoft Translator launches **Odia (Oriya)**, **Dari**, **Pashto** and **Sorani Kurdish** and **Kurmanji Kurdish**, bringing its total number of languages to 77.\n\n##### June 2020 - Microsoft Translator\nMicrosoft Translator launches **Kazakh**, bringing its total number of languages to 72.\n\n##### April 2020 - Microsoft Translator\nMicrosoft Translator launches **Marathi**, **Gujarati**, **Punjabi**, **Malayalam** and **Kannada**, bringing its total number of languages to 71.\n\n##### April 2020 - DeepL\nDeepL launches **Brazilian Portuguese** in addition to Portuguese, joining Microsoft Translator in offering two varieties as a target language. In contrast, Google's only option is Brazilian Portuguese.\n\n##### March 2020 - ModernMT\nModernMT announces **free machine translation** for April and May, unlimited other than by its hardware constraints.\n\n##### March 2020 - Yandex Translate\nYandex Translate launches **Yakut**, bringing its total number of languages supported to 97. However it's one of 4 languages not yet available in the API.\n\n##### March 2020 - DeepL\nDeepL adds **Chinese** and **Japanese**, its first non-European languages, bringing its total number of languages supported to 11. \n\n##### February 2020 - Google Translate\nGoogle Translate adds **Kinyarwanda**, **Odia (Oriya)**, **Tatar**, **Turkmen** and **Uyghur**, bringing its total number of languages supported to 108.\n\n##### January 2020 - Yandex Translate\nYandex Translate launches **Chuvash**.\n\n##### January 2020 - Microsoft Translator\nMicrosoft Translator launches **Irish Gaelic**.\n\n##### October 2019 - ModernMT\nModernMT launches **human-in-the-loop**, **document-level context** and **context adaptation**.\n\n---\n\n### Thanks\n\nWe're very thankful to all who have contributed to this comparison.\n\n","output":null,"collection":"landscapes","url":"/compare","relative_path":"_landscapes/compare.md","previous":null,"excerpt":"

This page was last updated on\nApril\n21st,\n2021. We generally compile this comparison only from public or publicly derivable information. You can send updates and suggestions to [email protected].

\n","next":{"collection":"landscapes","url":"/options","relative_path":"_landscapes/options.md","path":"_landscapes/options.md","id":"/options","draft":false,"categories":["landscape"],"title":"Options for translation quality and risk","description":"How to build or buy services, tools or technology for measuring translation quality and risk","layout":"default","tag":"translation-risk-prediction","category":"landscape","slug":"options","ext":".md","tags":["translation-risk-prediction"]},"path":"_landscapes/compare.md","id":"/compare","draft":false,"categories":["landscape"],"title":"Compare translation API features","description":"An overview of the features and limitations of the major machine translation APIs","layout":"default","image":"https://modelfront.com/assets/images/compare/compare.png","tag":"machine-translation","category":"landscape","slug":"compare","ext":".md","tags":["machine-translation"]},"excerpt":"

Is a translation good or bad? Measuring quality and risk are fundamental to successful translation at scale. Both human and machine translation benefit from sentence-level and corpus-level metrics.

\n","next":null,"path":"_landscapes/options.md","id":"/options","draft":false,"categories":["landscape"],"title":"Options for translation quality and risk","description":"How to build or buy services, tools or technology for measuring translation quality and risk","layout":"default","tag":"translation-risk-prediction","category":"landscape","slug":"options","ext":".md","tags":["translation-risk-prediction"]},"path":"_landscapes/compare.md","id":"/compare","draft":false,"categories":["landscape"],"title":"Compare translation API features","description":"An overview of the features and limitations of the major machine translation APIs","layout":"default","image":"https://modelfront.com/assets/images/compare/compare.png","tag":"machine-translation","category":"landscape","slug":"compare","ext":".md","tags":["machine-translation"]}
Languages Customization Context Pricing
for 1M chars
Data rights
Default terms for API calls
Google
Translate API
108 Terminology

AutoML,
to and from English only
$20
$80 for custom
No-trace
LingvaNex
Translate API
105 $5 No-trace
Yandex
Translate API
93 Terminology

Custom models
in private beta
$6
₽447
No-trace
Microsoft
Translator Text API
90 terminology

Custom Translator,
to and from English only,
limited pairs
$10
$40 for custom
No-trace
GDPR-compliant
HIPAA-compliant
Amazon
Translate API
73 Terminology

Adaptive customization,
to and from English only
$15
$60 for custom adaptive
It’s complicated.
ModernMT
API
47 Terminology

Adaptive customization
for all pairs
Document-level $15
$15 for custom adaptive
$50 for real-time adaptive
No-trace
GDPR-compliant
IBM
Watson Language Translator
39 Terminology

Custom models
$20
$100 for custom
Client-only
Stored temporarily
Baidu
翻译API
28 Terminology

Custom models

vertical-specific models
¥49
Custom unlisted
It’s complicated.
DeepL
API
24 €20 No-trace
Alibaba
Machine Translation
15
Limited pairs
Commerce-specific model $33 Client-only
Stored temporarily

Only features available via API are considered. For example, Google supports gender variants, and DeepL supports context awareness across lines, but those features are only available in their consumer websites and applications, not via their APIs.

Language counts consider the listed languages, no matter how similar, and assume essentially all pairs of those languages are supported unless otherwise noted.

Pricing considers the paid rate at moderate volume, not free quota or volume discounts, nor prepaid or subscription discounts. Most of the APIs offer free quota, and charge small additional fees for training custom models and keeping them deployed. There is also continued widespread unpaid use of unofficial APIs based on reverse-engineering consumer translation applications.

Some systems use different architectures and even approaches for different language pairs or content types.

For more details, please see the documentation of each API.


Other machine translation APIs

There are other self-serve APIs to consider, notably those from Naver, eBay, Tencent and Youdao. There are also APIs like the Intento and the Rakuten RapidAPI that aggregate many translation APIs and more.

Other machine translation features

There are other features to consider, like locale, formality, data privacy, language identification, HTML and template translation, input correction, transliteration, speech input, latency, batching, offline models and on-prem deployment, that are out of the scope of this comparison for now.

For example, Google Translate produces French mostly with formal vous, Parisian French word choices but Canadian French punctuation. Microsoft Translator now offers a choice between Parisian and Canadian French, but if you want informal French with tu out of the box, you'll have to make some compromises - none of the major APIs produces it consistently.

Other machine translation providers

SYSTRAN, Unbabel, Lilt, Tilde, Omniscien, KantanMT, PangeaMT, TextShuttle, Globalese, Tarjama, Iconic, AppTek, SAP and more train custom machine translation, but do not offer a self-serve API with public per-character pricing. SYSTRAN offers an open-source library, OpenNMT.

Other machine translation players

Apple, Facebook and Rakuten have their own machine translation deployed in their own consumer platforms, but do not offer an official API. Facebook does offer an open-source library, Fairseq, and pre-trained models.

To the best our knowledge at this time, many others with machine translation in their platforms, like Twitter and AirBnB, as well as translation providers and CAT tools like Lionbridge and SDL, use the APIs listed above or on-premise deployments of other providers and do not develop their own machine translation from scratch.

Machine translation libraries

Many major machine translation providers also offer open-source libraries and even pre-trained models.

Owners Users Framework Language
Fairseq Facebook Facebook PyTorch Python
Marian NMT Microsoft Microsoft Translator, Unbabel, Rakuten, eBay, Tilde, Project Bergamot, Opus-MT C++
ModernMT
based on Fairseq
ModernMT ModernMT PyTorch Java, Python
OpenNMT SYSTRAN, Ubiqus SYSTRAN, Ubiqus, Tarjama Tensorflow, Pytorch Python
Trax Transformer Google Google Translate Tensorflow Python
Sockeye Amazon Amazon Translate MXNet Python

Note that there are many more users of all these libraries and clients of the APIs that use them. It is not possible to track all closed use of open-source code. For example, ModernMT is used by Translated, AirBnb and the European Parliament.

Pre-trained machine translation models

Opus-MT offers pretrained Marian NMT models for many language pairs.

Facebook offers pre-trained Fairseq models for a limited number of languages.

They are also available via Hugging Face.


Updates

March 2021 - DeepL

DeepL launches Bulgarian, Czech, Danish, Estonian, Finnish, Greek, Hungarian, Latvian, Lithuanian, Romanian, Slovak, Slovenian, and Swedish, bringing its total number of languages to 24.

February 2021 - Microsoft Translator

Microsoft Translator launches Albanian, Amharic, Armenian, Azeri Turkish, Khmer, Lao, Myanmar, Nepali, and Tigrinya, bringing its total number of languages to 90. It is the first API to support Tigrinya.

January 2021 - Microsoft Translator

Microsoft Translator launches Inuktitut first, bringing its total number of languages to 79.

November 2020 - Amazon Translate

Amazon Translate launches adaptive translation as Active Custom Translation, joining ModernMT, and launches more than a dozen new languages, bringing its total number of languages to 73.

October 2020 - Microsoft Translator

Microsoft Translator launches support for Canadian French.

September 2020 - Microsoft Translator

Microsoft Translator launches Assamese first, bringing its total number of languages to 78. It also updated its system for training custom models.

August 2020 - LingvaNex

LingvaNex launches their translation API with more than 100 languages, second only to Google Translate.

August 2020 - Microsoft Translator

Microsoft Translator launches Odia (Oriya), Dari, Pashto and Sorani Kurdish and Kurmanji Kurdish, bringing its total number of languages to 77.

June 2020 - Microsoft Translator

Microsoft Translator launches Kazakh, bringing its total number of languages to 72.

April 2020 - Microsoft Translator

Microsoft Translator launches Marathi, Gujarati, Punjabi, Malayalam and Kannada, bringing its total number of languages to 71.

April 2020 - DeepL

DeepL launches Brazilian Portuguese in addition to Portuguese, joining Microsoft Translator in offering two varieties as a target language. In contrast, Google's only option is Brazilian Portuguese.

March 2020 - ModernMT

ModernMT announces free machine translation for April and May, unlimited other than by its hardware constraints.

March 2020 - Yandex Translate

Yandex Translate launches Yakut, bringing its total number of languages supported to 97. However it's one of 4 languages not yet available in the API.

March 2020 - DeepL

DeepL adds Chinese and Japanese, its first non-European languages, bringing its total number of languages supported to 11.

February 2020 - Google Translate

Google Translate adds Kinyarwanda, Odia (Oriya), Tatar, Turkmen and Uyghur, bringing its total number of languages supported to 108.

January 2020 - Yandex Translate

Yandex Translate launches Chuvash.

January 2020 - Microsoft Translator

Microsoft Translator launches Irish Gaelic.

October 2019 - ModernMT

ModernMT launches human-in-the-loop, document-level context and context adaptation.


Thanks

We're very thankful to all who have contributed to this comparison.