Translated Launches Lara 3, a Translation AI That Learns From Its Own Mistakes

Translated has introduced Lara 3, the third generation of its Lara Translate AI, built on a training method the company calls learning by doing. The release is backed by 27 years of machine translation experience at Translated and represents what the company describes as the largest quality jump ever measured between two Lara generations.

Breaking the Imitation Ceiling

Machine translation has moved through several technical eras, from rule-based systems to statistical translation to neural networks and, more recently, large language models. Despite the different underlying math, each generation shared the same limitation: models learned by imitating human reference translations, meaning quality was capped by whatever was in the training data. No dataset contains a perfect translation of every sentence in every context, so imitation alone could only go so far.

Lara 3 adds a step earlier systems lacked. During training, the model produces multiple variations of a translation, an automatic judge built from professional reviewer expertise scores each attempt against professional standards, and the model refines its output based on that feedback. Run across millions of sentences, that loop of exploration, judgment, and refinement lets the model learn what quality actually looks like in context, rather than memorizing a single reference answer per sentence.

Topping the Benchmarks

In blind human evaluations on the WMT2025 benchmark, which spans books, news, and real conversational text, Lara 3 finished first, ahead of frontier models including Fable-5 and GPT-5.6 Sol, and clearly ahead of Google and DeepL. Because the evaluations were blind, reviewers scored output without knowing which system produced it, so the result reflects quality rather than brand recognition. Lara 3 also led every domain tested across enterprise localization benchmarks in travel, technology, and finance, spanning all 21 language pairs evaluated.

Those quality gains came without a speed or cost trade-off. Compared with Fable-5, Lara 3 runs its translation workflow about 23 times faster while delivering close to four times the translation capacity for the same budget.

A Fully Multimodal Platform

Lara 3 extends well past text. The release adds image translation that preserves original layout while localizing small, angled, and shaded text, built directly into the mobile app, and audio translation that keeps the original speaker's timbre while translation memories and glossaries shape accuracy. Document translation now covers 72 file formats with a 70% reduction in layout errors and automatic font resizing to prevent broken pages, and a browser extension lets users translate entire websites in place as they browse.

Developer-facing tools round out the release: Lara Think, a reasoning mode that cuts errors nearly in half on difficult passages; Lara Prosa, tuned for literary and editorial content and built to hold coherence across long context; multilingual profanity detection and filtering; a command-line interface; and an MCP server that connects Lara directly into assistants such as ChatGPT, Claude, and Groq.

Pricing and Availability

Lara 3 moves away from per-seat and per-token billing models, both of which penalize teams for adding colleagues or feeding the model useful context. Organizations instead pre-purchase a shared character allowance that unlimited team members can draw from, with charges applied only to source characters translated. Glossaries, context, translation memory, and instructions carry no additional cost, and extra usage is available on demand.

Lara 3 also gives customers a choice in how their data is handled, balancing data protection against model improvement, with data residency available in either the European Union or the United States, addressing requirements common among regulated industries.

Lara 3 is live for selected partners now, with public availability planned in the coming weeks. The current version of Lara Translate is available at laratranslate.com.