Translated Launches Lara 3, Making the Case for Specialized Translation AI

A translation can read beautifully and still get the job wrong. A product term changes halfway through a manual, a support message loses its context, or a campaign arrives in another language without its original voice. Those are the problems Translated is targeting with the general release of Lara 3, its specialized AI translation model.

Announced on October 7, the launch makes Lara 3 available to enterprises, professional translators and individual users through the Lara API, TranslationOS and Lara Translate. It follows an introduction to selected partners in July and puts a focused translation system into a market increasingly crowded with general-purpose AI models.

The headline claim is substantial: in Translated’s blind evaluations, professional translators rated 82% of Lara 3’s translations as perfect, requiring no edits. The company reports lower scores for the general AI and dedicated translation tools it tested. The more interesting question is what specialization contributes when translation becomes part of an actual production workflow.

What the benchmark measures

According to the launch release, the evaluation covered 16,800 translations from English into 21 enterprise localization languages. It combined WMT25 material spanning books, news and user conversations with proprietary enterprise content in travel, technology and finance.

Averaged across the 21 language pairs, Translated reports the following shares of translations rated perfect:

• Lara 3: 82%
• GPT-5.6 Sol: 69%
• Claude Fable 5: 67%
• DeepL: 57%
• Google Translate: 53%

The release says Lara finished first in 20 of the 21 language pairs. Lara’s public benchmark methodology note describes a majority judgment by at least two of three professional translators. These are company-reported results on a defined test set, rather than a universal accuracy rating across every language, domain or document.

That distinction matters. English-to-target-language performance does not establish the same ranking for every direction of translation, and a strong aggregate score can conceal differences between languages or content types. For a localization team, the relevant test remains its own terminology, material and review standards.

A model that practices before deployment

Translated’s technical introduction to Lara 3 describes a training approach it calls “learning by doing.” Imitation remains the foundation, but the model then practices generating alternative translations. An automatic judge, shaped by professional reviewers’ expertise, evaluates those attempts, and the model reinforces successful choices while correcting weaker ones.

The company says this cycle runs millions of times during training. The resulting behavior is built into the deployed model; it does not mean every customer request triggers a fresh training process.

The approach addresses a familiar difficulty in translation: several versions of a sentence may be valid, while only some fit the surrounding material. Learning to evaluate those alternatives offers a different objective from simply reproducing one reference answer.

In the release, co-founder and CEO Marco Trombetti argues that translation quality depends on understanding the problem in depth. Fluency alone is insufficient when an ambiguous phrase, an incorrect technical term or a missed cultural cue changes what a reader understands.

Context is part of the workflow

Training is only one layer. Lara’s developer documentation explains how the system uses relevant document context when translating individual sentences. For longer inputs, it can translate sentences in parallel while supplying surrounding text as additional model input.

Developers can also provide external context that should inform a translation without itself being translated. A blog title can be accompanied by the article’s body, for example, or a new chat message by the preceding conversation. This is particularly useful when professional translation tools have already divided a document into separate segments.

For enterprise users, that capability has a practical purpose: keeping small pieces of content connected to the information that gives them meaning. Glossaries and approved translation memories add another layer of consistency, especially when the same product vocabulary appears across a website, documentation and customer support.

From quality scores to production economics

The release also reports separate speed and cost benchmarks. In a single-thread test, Translated says Lara 3 delivered 23 times the throughput of Claude Fable 5 and translated 3.75 times as many characters for the same budget. Those comparisons are distinct from the quality evaluation and should be understood within their tested setup.

Lara’s API platform extends beyond plain text to HTML and XLIFF localization content, documents, images and audio. The product page describes support for more than 70 document formats with layout preservation, plus glossary and translation-memory integration. Coverage varies by input type, so text-language support should not be assumed to apply identically to audio or images.

Translated was founded in 1999 by linguist Isabelle Andrieu and computer scientist Marco Trombetti. Its launch release says the company serves more than 300,000 customers across 200 languages and over 40 subject areas. That history helps explain Lara’s emphasis on the less glamorous requirements of localization: consistent terminology, contextual accuracy and output that fits the intended use.

Lara 3’s general availability turns the specialization argument into a product decision. Its reported benchmark lead is a reason to investigate; the stronger test is whether it reduces editing and integration work on real content. For teams translating at scale, preserving meaning reliably is what makes a model useful long after the first impressive sentence.