Models

The AI models and MT engines evaluated: their features, language coverage and results.

Takeaways for , , , and

Takeaway 1

We benchmark 57 models from 38 vendors: 28 translation APIs and 29 LLMs prompted for translation.

Eight of the translation APIs are LLM-based services, and six of the LLMs are open-weight models whose weights anyone can download. We aim to include every commercially available MT engine and LLM, and we use each one through a paid account, as a customer would. We use Intento Language Hub for uniform API access to MT engines and LLMs, with deterministic request routing, request batching, quota and account management.

The list holds the models available on 16 July 2026, when we froze it. The last ones added before the freeze were Claude Fable 5, Claude Sonnet 5 and GPT-5.6. Meta Llama 4 Maverick joined on 1 September as the only exception, and it runs on the FP8 build its inference provider serves. We also benchmark GPT-4.1, which is still widely deployed in production, next to the GPT-5 models, so a reader can decide on evidence whether to migrate from it.

We set up API access to 64 models, and seven of them did not return translations of our test texts, so they are not in the results: Amazon Nova 2 Pro and Nova 2 Lite, EuroLLM-22B, Sarvam-Translate, T-tact AN-ZIN, TREBE and Yaraku. Five vendors give access to their models only after approval: Mirai, Kakao and GPT-NL declined, and Rozetta and Tilmoch did not reply. Translated launched a new version of Lara after the model freeze, so the results describe Lara 2.

See details →

Vendors of the 57 models in the benchmark

Takeaway 2

We classify models by API access, distinguishing MT engines from promptable LLMs.

We call an "MT engine" a model that is accessible via a translation API, and an "LLM" a model with access to prompting. Within these two classes we mark two subclasses from vendor information: an "LLM-based MT engine" is an MT engine built on an LLM, such as DeepL Next Gen and Google Translation LLM, and an "open-weight LLM" is an LLM with downloadable weights, subject to its license terms.

See details →

Takeaway 3

The best MT engine translation has more errors than the best LLM translation in all 80 languages with results for both.

The median ratio is ~2.3 times as many errors in the best MT engine translation.

The best MT engine is DeepL Next Gen in 49 languages and Translated Lara 2 in 23. The gap is smallest in Romanian, European Portuguese and Bosnian, and only the DeepL Next Gen solutions in German and Romanian meet general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words).

See details →

Errors in the best MT engine and LLM solutions, by language

Takeaway 4

The same LLMs have the fewest general translation quality errors in nearly every language on their shortlists.

Translations from OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5, OpenAI GPT-5.5, GPT-5.6 Terra and Google Gemini 3.1 Pro Preview have fewer errors than the median of the other shortlisted LLMs in 95% to 100% of the languages where we shortlisted them. For OpenAI GPT-4.1, Google Gemma 4 and Mistral Large 3, the share is under 15%. The number of errors depends on the language as well, but the order of the models changes little from one language to another.

See details →

Languages where each LLM has fewer errors than the other shortlisted LLMs at the median

Takeaway 5

LLMs account for 99% of one-step solutions that meet the Basic requirements.

The LLM share among the top performers was 89% in 2025 and 55% in 2024, using the earlier definition described below.

The other 1% come from DeepL Next Gen, an LLM-based MT engine, in German and Romanian. We shortlist at least three real-time solutions in each language, nearly all of them MT engines, so MT engines make up 23% of the one-step solutions we assess with Intento LQA.

In the previous edition, a top performer was a solution whose quality score was statistically tied with the best one in its language pair, so the earlier shares come from a different way of counting.

LLMs among the best translation solutions, by edition

Takeaway 6

Critical errors occur more often in MT engine translations than in LLM translations.

No critical error occurs in 56% of MT engine solutions and 89% of LLM solutions.

Major and critical errors make up 49% of the weighted errors in MT engine translations and 30% in LLM translations. The gap holds at the same number of errors. With 35 to 50 errors per 1,000 words, 42% of the MT engine solutions have no critical error, against 80% of the LLM solutions.

Translations from Google Translation LLM, Amazon Translate and ModernMT Enterprise Edition have a critical error in every language where we shortlisted them.

See details →

Translation errors per 1,000 words by severity, by model type

Takeaway 7

LLM-based MT engine translations have ~half as many errors as other MT engine translations, but rarely meet formatting requirements.

They have ~1.6 times as many errors as LLM translations in the same language, at the median. The LLM-based MT engines on our shortlists are DeepL Next Gen, Translated Lara 2, Google Translation LLM and Widn AI Tower Sugarloaf, and the other MT engines are mostly Microsoft Translator.

1% of the LLM-based MT engine solutions meet the formatting requirement, against 84% of the LLM solutions.

Weighted translation errors per 1,000 words, by model type

Takeaway 8

We verify language support by checking whether each model returns a usable translation in that language.

Vendor documentation does not reliably say which languages a model supports. LLMs have no language list, and when we tested the language lists of MT vendors, they were wrong in both directions: we did not get a translation in some listed languages, and got translations in some unlisted ones.

From 21 of the 57 models, in 94 of the 100 languages, we got text in the wrong language, garbage or a near-empty output, a translation cut off part-way, text in the wrong script, or text far shorter than the source. We remove such outputs before scoring quality, because a quality score for a text that is not a translation means nothing. A model supports a language only when we got a translation in it from the model, and that is how we determine language support for every model in the report.

See details → · Explore interactively →

Language support by model

Takeaway 9

The strongest shortlisted LLMs meet our language-support criteria in nearly every language assessed, and the weakest in at most half.

We count a language as supported by an LLM when the translation of our four texts has at most one critical error, no hallucinated segment and at most 60 errors per 1,000 words, weighted by severity. A language where we get no translation from the LLM at all counts as not supported.

Translations from Anthropic Claude Fable 5 and OpenAI GPT-5.6 Sol meet these criteria in every language where they are on the shortlist, and from Google Gemini 3.1 Pro Preview in all but Burmese. Translations from OpenAI gpt-oss-120b and xAI Grok 4.20 (non-reasoning) meet them in 6 of 12 and 3 of 8 languages.

Translations from Fable 5 meet the much higher bar of general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words) in 33 of the 60 languages. We assess an LLM in a language only where it is on our shortlist.

See details →

Measured language support per LLM in 100 languages

Takeaway 10

We exclude unusable translations before scoring quality.

We excluded translations from five models in 15 to 68 languages each, and from 15 more models in one to 13 languages each.

We get one or two translations from a model in each language, one without context and one with context in the request. We exclude each one on its own. Of the 235 model and language pairs with an exclusion, we excluded every translation in 134, only the one without context in 51 and only the one with context in 50.

The most frequent reasons are garbage or repetition loops and text in the wrong language, in 95 pairs each, and cut-off text, in 86, while the error messages in place of the translation are from two models. With context in the request, we got text in the wrong language or script less often: where we have both translations, only the one without context had it in 37 pairs and only the one with context in 10, while the difference for cut-off or garbled text (53 and 39 pairs) is small enough to be chance.

We excluded at least one translation in 24% of the model and language pairs of open-weight LLMs and in 11% of those of LLM-based MT engines, against 3% for other LLMs and 2% for MT engines. From the remaining 37 of the 57 models, every translation is in the results.

See details →

Takeaway 11

We use zero data retention wherever the model API supports it.

Zero data retention means the vendor does not store the requests sent to its API, under the applicable retention terms. For enterprise translation this matters because the content is often confidential, and the texts in this benchmark are enterprise content, contributed on our promise to use them for evaluation only, never for training.

The API of Claude Fable 5 returns an error unless the account has data retention turned on, so we translate with it in a dedicated workspace. The workspace has its own API key and keeps requests for 30 days, and we use it only for the benchmark texts.

More on Models

Takeaway 0

The best open-weight LLM translation has more errors than the best other LLM translation in all 50 languages with results for both.

The median ratio is ~2.9 times as many errors in the best open-weight LLM translation.

The gap is smallest in Indonesian, where Google Gemma 4 has the only open-weight solution that meets general translation quality (no critical error, at most three major errors and at most 20 weighted errors per 1,000 words), and it meets every other Standard requirement too. The open-weight LLMs on our shortlists are Google Gemma 4, OpenAI gpt-oss-120b, Meta Llama 4 Maverick, Alibaba Qwen3-235B-A22B and Swiss AI Apertus-8B.

See details →

Errors in the best open-weight and other LLM solutions, by language

Takeaway 0

Some translation errors and unmet requirements recur across languages for the same model.

The translated JSON document is invalid in every language for Translated Lara 2, Microsoft Translator, Google Translation LLM and Widn AI Tower Sugarloaf. Translations of the JSON document from Anthropic Claude Opus 4.8 meet the tag handling requirement in 12 of 46 languages, against 73 of 76 languages for Google Gemini 3.1 Pro Preview in every document.

Translations from Translated Lara 2 carry an HTML code in place of each newline character between the tags and the JSON fields of the source file, in every language. Without these codes, the translated JSON document would be valid in 36 of the 44 languages. In nearly every language, translations from Google Translation LLM carry additions (sentences that are not in the source).

We translate every text with its HTML tags and set the format parameter for tagged content where the API has one, textType html for Microsoft Translator, mimeType text/html for Google Translation LLM and content_type text/html for Translated Lara 2. We don't set a format parameter for Widn AI Tower Sugarloaf.

See details →

Share of one-step solutions meeting the tag handling and formatting requirements, by model

Takeaway 0

The vendor list grew from 21 in December 2018 to 106 in August 2026, mainly through LLM providers after 2023.

Since then the number of MT vendors with a standalone API has changed little (from 48 to 57). Six vendors sell LLM-based translation models (DeepL, Translated, Tilde, NiuTrans, Youdao and TransPerfect).

MT vendors and LLM providers both change fast. Within a year, 18 of the 51 MT vendors of 2025 left the market, changed name or owner, or are non-commercial and out of our scope, and 19 of the 41 LLM providers of 2025 now have only outdated or English-only models. Without those 19, the total is 85. If you chose a default model a year ago, its provider may no longer release multilingual models.

See details →

Vendors of machine translation and LLMs, December 2018 to August 2026

Takeaway 0

Of 59 machine translation vendors, 31 support custom terminology.

Custom terminology support means you can apply your own glossary through the API, in training or at translation time. Fewer vendors support domain adaptation. 21 offer static domain adaptation, a batch retraining on thousands of your parallel segments that takes hours to days. 10 offer dynamic domain adaptation, which updates the model in real time from as few as one example. 20 vendors offer an LLM-based translation model, which is normally an LLM fine-tuned for translation and available through an MT API, without prompting.

See details →

Machine translation vendors, August 2026

Takeaway 0

Of 33 vendors with promptable LLMs, 25 offer open weights for commercial use.

A vendor with a model you can prompt freely is on the generative AI chart, and a vendor with a model that only translates is on the machine translation chart. A vendor can be on both: Cohere sells Command A Translate next to Command A. For DeepL, Translated, NetEase Youdao, ALL-Lab, NiuTrans and Xiaomi we track only translation models, so they are on the machine translation chart only.

We group the providers by license class: cloud commercial, open commercial, open conditional and open non-commercial. Open conditional weights are free to download and, in principle, to use commercially, gated by a scale threshold, an acceptable-use policy or a registration step. 26 providers offer a flash tier, a low-latency and lower-cost model line, and we leave out 19 that have only outdated or English-only models.

See details →

Generative AI providers by license class, August 2026

Takeaway 0

Vendor rules for access, billing, region and request size can affect translation cost.

The rules fall into four kinds: access (keys, trials, licenses), billing and quota (prepaid balances, character quotas, minimum charges), region (endpoints and resources bound to one region) and request limits (parameters and request sizes). Some vendors have a minimum charge per request (5,000-20,000 characters), which raises the cost of short texts. During this benchmark we use Intento Language Hub, which provides a single API to all NMT and LLM models and handles differences in how the APIs express translation requests.

Takeaway 0

We standardize language codes before comparing language support or making translation requests.

Simplified Chinese is zh at Amazon, zh-CN at Google, zh-Hans at Microsoft, zh-CHS at Youdao and cmn-Hans-CN at Tencent. Baidu writes Japanese as jp and French as fra, and Translated Lara 2 indexes Yiddish as ydd-US. A language listed under an unexpected code reads as not supported, and a code mismatch alone can drop a model from a comparison. We use ISO 639-1 two-letter codes, the ISO 639-3 code where a language has no two-letter one (ceb for Cebuano), and a script subtag for a language with two scripts (zh-hant, sr-latn).

See details →

Takeaway 0

We use temperature 0 for 15 of the 29 LLMs and document the settings for the other 14.

We can get different translations of the same text from one request to the next, because generation in an LLM includes randomness. Temperature is the API parameter that sets the degree of this randomness, and we use temperature zero to reduce variation where the API supports it. We want it at zero in a benchmark, for comparable output from the same input.

Three models of one vendor reject the temperature parameter with an API error, five reasoning models of two other vendors do not accept it at our settings, and one open-weight model has no temperature setting in its API, so these nine translate at vendor defaults. One model requires temperature 1, and we set four open-weight models to 0.1, the default of our connector for them.

We set the reasoning effort low on six models and turned extended thinking off on one. No request carried a fixed random seed, so we keep the archived outputs as the only exact record of each translation.

According to the Gemini 3 developer guide, as of August 2026, temperature should stay at its default of 1.0, because lower values can degrade reasoning. Only in the Cohere documentation is temperature 0 still fully deterministic.

See details →

Models

Takeaways for Localization managers, Business leaders, Linguists, and other rolesLocalization engineersLSPsTech vendors (TMS / MT)AI labs / model buildersAnalysts / pressProcurementAcademics──────────See everything

We benchmark 57 models from 38 vendors: 28 translation APIs and 29 LLMs prompted for translation.

We classify models by API access, distinguishing MT engines from promptable LLMs.

The best MT engine translation has more errors than the best LLM translation in all 80 languages with results for both.

The same LLMs have the fewest general translation quality errors in nearly every language on their shortlists.

LLMs account for 99% of one-step solutions that meet the Basic requirements.

Critical errors occur more often in MT engine translations than in LLM translations.

LLM-based MT engine translations have ~half as many errors as other MT engine translations, but rarely meet formatting requirements.

We verify language support by checking whether each model returns a usable translation in that language.

The strongest shortlisted LLMs meet our language-support criteria in nearly every language assessed, and the weakest in at most half.

We exclude unusable translations before scoring quality.

We use zero data retention wherever the model API supports it.

More on Models

The best open-weight LLM translation has more errors than the best other LLM translation in all 50 languages with results for both.

Some translation errors and unmet requirements recur across languages for the same model.

The vendor list grew from 21 in December 2018 to 106 in August 2026, mainly through LLM providers after 2023.

Of 59 machine translation vendors, 31 support custom terminology.

Of 33 vendors with promptable LLMs, 25 offer open weights for commercial use.

Vendor rules for access, billing, region and request size can affect translation cost.

We standardize language codes before comparing language support or making translation requests.

We use temperature 0 for 15 of the 29 LLMs and document the settings for the other 14.

Automate translation to meet your requirements

Takeaways for , , , and