Introducing Lara 3

AI enters the age of experience

In summary

Translated introduced Lara 3, its third-generation AI model for translation, with a new “Learning by Doing” training phase.

The release combines higher translation quality with major gains in speed, cost efficiency, and enterprise usability.

Lara 3 led the public WMT2025 benchmark and an enterprise benchmark covering 21 language pairs.
It delivers workflows over 23× faster and nearly 4× more translation capacity than Fable 5.

New capabilities are available in the Lara Translate platform, including image, audio, rich-text, and document translation, as well as Lara Think for extreme precision and Lara Prosa for an editorial layer when style matters as much as accuracy.

Pricing now uses shared source-character allowances, with unlimited colleagues and EU or U.S. data residency.

Read more to explore Lara 3’s training approach, benchmark results, features, and pricing.

ROME, July 30, 2026

Lara 3 represents a fundamental shift in how translation AI is built and delivered. Backed by our 27 years of experience in machine translation, the third generation of Lara moves beyond imitation-based training and introduces “learning by doing,” a training paradigm in which the model practices translation, receives feedback rooted in professional reviewer expertise, and refines its own output before ever serving a customer. According to blind human evaluations, the result is the largest quality leap ever observed between two Lara generations, achieved alongside dramatic improvements in speed and cost efficiency. The release also brings a broad set of new product capabilities, spanning image, audio, and document translation, and a redesigned pricing model built around transparency and organization-wide access.

From Imitation to Experience

The history of machine translation has unfolded through distinct technological eras. Rule-based systems, in which humans manually encoded linguistic rules, dominated from 1999 to 2005. In 2006, statistical machine translation marked the first breakthrough of machine learning in the field: rules were learned from data rather than written by hand. In 2014, sequence-to-sequence architectures brought neural networks to translation, and in 2017 the transformer, introduced in the landmark paper “Attention Is All You Need”, enabled models to leverage long context for dramatically better output. In 2022, the industry entered the era of LLM-based translation, specializing large language models for the translation task.

Every one of these generations, however, shared a common foundation: imitation. Models were shown a source sentence and a reference translation and trained to replicate that reference. The approach works, but it has an inherent ceiling. For any given sentence, multiple correct translations exist; training a model to match only one of them teaches it to reproduce a translation, not to understand why a translation is good. A model trained purely by imitation can, by definition, only be as good as its training data.

Lara 3 is designed to break that ceiling. Think of how humans learn: children learn first by imitation, then by pursuing goals set by parents and teachers, and finally, as adults, by self-judgment, evaluating their own work and improving without external instruction.

It is this final capacity, self-judgment, that Lara 3 introduces to machine translation.

How Learning by Doing Works

Imitation remains the foundation of training, but Lara 3 adds a final phase in which the model enters a virtual practicing environment.

Three things happen there. First, Lara explores: given a sentence, it produces multiple variations, experimenting with style, tone, and wording. Second, an automatic judge evaluates each attempt, identifying what works and what doesn’t, down to whether a specific word is correct in a specific context. Third, Lara refines: it incorporates that feedback into its learning, reinforcing hypotheses that score well and correcting those marked as errors. This cycle repeats millions of times, and it all happens during training, by the time Lara serves its first translation, the learning is already baked into the model.

The distinguishing element is the judge itself, the part that only we could build. Rather than scoring translations against abstract or generic evaluation metrics, it was created from the deep expertise of professional reviewers working on real translations. Their standards, their reasoning, and their corrections shape the feedback Lara practices against, the same type of feedback professional translators give in their day-to-day work. For enterprise teams, this means the model’s standards are effectively your standards.

Benchmark Performance

In blind human evaluations, Lara 3 tops the public WMT2025 benchmark, which covers books, news, and user conversation, edging out frontier models including Fable-5 and GPT-5.6 Sol and clearly surpassing Google and DeepL. Enterprise localization, however, is considerably harder than general-purpose translation: teams demand consistent brand voice, terminology and style guide compliance, product context, and content that feels native in every market. On a production benchmark built specifically for enterprise localization, Lara 3 leads in every domain tested, travel, technology, and finance, and in all 21 language pairs evaluated.

Quality is only one part of the production equation. Compared with Fable-5, Lara delivers the same workflow more than 23 times faster, a difference that makes real-time integration practical: product content, support experiences, and customer-facing journeys can be translated without noticeable delay. On cost, Lara provides almost four times more translation capacity than Fable-5 for the same budget, making it the most cost-efficient option among the competitors evaluated. Taken together, teams no longer need to trade off between quality, speed, and cost.

New Product Capabilities

Lara 3 ships with an extensive set of features, available immediately. Image translation preserves layout with striking fidelity, small text, angles, and shading are all localized, and is integrated into the mobile app, where you can photograph a restaurant menu or any document and receive a translated version.

Audio translation preserves the original speaker’s timbre while applying the same adaptive quality of Lara’s text engine, including translation memories, glossaries, and style guides.

For text workflows, you can now copy and paste rich content from any application and retain full formatting in the web interface.

Document translation, already supporting 72 file formats, sees a 70% reduction in layout errors, with automatic font resizing to keep layouts intact when translations run longer than the source.

A new browser extension brings website and in-browser translation to any environment.

Developers gain several additions: Lara Think, a reasoning mode that cuts errors nearly in half when higher quality is required; Lara Prosa, designed for literature and editorial content with long-context coherence and publication-specific style; multilingual profanity detection and filtering; and a command-line interface plus an MCP server that make Lara fully agentic, integrating with ChatGPT, Claude, and Groq.

Transparent, Flexible Pricing

Lara 3 abandons per-seat and per-token billing. Token-based pricing is inherently unpredictable: providers define what a token is, and reasoning models can generate 10 or 100 times more output than expected. Lara instead charges only for source characters, no charges for glossaries, context, translation memory, or instructions.

Your organization pre-purchases a character allowance shared by everyone: a single account can invite unlimited colleagues, with all features (text, documents, images, audio) billed against the same pool, and extra usage available on demand. If you know your volume, you know exactly what you will spend.

Privacy, Security, and Availability

Customers can balance data protection against model improvement through multiple privacy options, rely on certifications designed to satisfy InfoSec requirements, and choose data residency in the European Union or the United States. Lara 3 is available to selected partners today and will be publicly available in the coming weeks; product features can be explored now at laratranslate.com.