Translated combines proprietary AI technology with a network of more than 400,000 professional translators to deliver translation services across more than 200 languages. Its technologies include TranslationOS, Lara, ModernMT, T-Rank, Matecat, MateSub, and MateDub. Pulse 2.0 interviewed Translated Co-Founder Marco Trombetti to learn more.
Description Of The Company
When asked to describe Translated and its mission, Trombetti explained:
My wife Isabelle and I started Translated in 1999 because we believed AI could help professional translators work better, not replace them. That distinction still matters. Today, we serve over 500,000 customers across 200 languages, including Airbnb, SpaceX, and Uber. But size was never the goal. The goal is allowing everyone to understand and be understood in their own language.
What makes us unusual is that we’ve always approached translation as a research problem, not just a services business. We build our own machine translation technology. That includes Lara, our translation AI that now outperforms popular systems like Google Translate and DeepL, and ModernMT, our adaptive neural engine. We also maintain a network of over 400,000 professional translators and invest heavily in research. Right now, we’re coordinating DVPS, a €29 million European Union project on multimodal AI with 20 institutions, including Oxford, ETH Zurich, and the Alan Turing Institute.
We also pursue partnerships that show how translation serves larger missions. This year, we are working in partnership with the Vatican to provide real-time translations of Mass at St. Peter’s Basilica in 60 languages, which will enable millions of visitors to follow the liturgy in their own language. That combination of scale, research depth, and meaningful partnerships is rare in our industry.
How Translated Started
When discussing what led him and Isabelle to create Translated, Trombetti shared:
I’m a computer scientist; Isabelle is a linguist. In the late 1990s, we watched the internet change how information moved across borders. It was obvious that language would become a bottleneck. Machine translation at the time was terrible, but the core insight seemed right: combine machine intelligence with human expertise, and you might get something better than either alone.
The other thing that drove me was seeing how language barriers slow progress. Scientific knowledge trapped in one language. Business opportunities missed because of translation costs. People unable to access information in their own language. These are civilizational problems. And I’m an optimist. If a problem seems that important, I’d rather spend my life working on it than assume someone else will.
Core Products And Features
When asked about Translated’s core products and features, Trombetti detailed:
Our core offering is professional translation delivered through TranslationOS, which is how large companies manage multilingual content. It integrates three proprietary technologies. First, Lara, our translation AI that now supports over 200 languages. Second, ModernMT, our adaptive neural machine translation technology that learns in real time from translator corrections. Third, T-Rank, which matches each project to the best translator from our network based on over 30 factors, such as domain expertise and past performance.
For translators, we built Matecat, an open-source computer-assisted translation tool, plus MateSub and MateDub for multimedia work. For consumers, we launched Lara Translate, giving individuals access to professional-grade translation AI.
The Vatican partnership is a good example of how these technologies work together. We provide real-time translations of Mass at St. Peter’s Basilica in 60 languages. Pilgrims scan a QR code and follow the liturgy on their smartphones. It’s technically sophisticated but invisible to the user, which is what technology should be. Cardinal Gambetti said it well: the tool helps serve the mission that defines the Catholic Church, universal by its very vocation.
The principle is the same across all our products: human and machine working together. The machine handles repetition and speed. The human handles judgment and nuance. Neither works well alone.
Translation Industry Challenges
When asked about the most prominent challenges facing the translation sector and how Translated is preparing for them, Trombetti noted:
The challenge isn’t technological anymore. Lara already approaches the quality of top-tier professional translators. With our next model, we expect to reach what we call “language singularity,” where checking a machine translation takes a professional translator no longer than checking a colleague’s work. We’re almost there.
The real challenge is application. How do we deploy this where it matters most? That requires partnerships with organizations that serve large, diverse populations. The Vatican collaboration is a good example. St. Peter’s welcomes millions of pilgrims annually who speak hundreds of languages. Providing real-time liturgical translation in 60 languages isn’t just technically complex; you need to understand the context, the acoustics, and the sacred nature of the content.
The second and most pivotal challenge is ensuring these tools serve people rather than displace them. I’m not interested in automating translators out of work. I want them handling higher-value work while machines handle repetition, enabling humans to focus on the most important tasks. That requires designing systems where humans remain essential to the architecture.
Then there are obvious challenges in low-resource languages. AI has mostly benefited languages with massive datasets. But there are thousands of languages spoken by smaller communities. This is where DVPS comes in. We’re building AI systems that learn from direct interaction with the physical world, combining visual input, spatial audio, speech direction, and sensor data. Multimodal models can learn from less textual data because they have signals from vision and context. This will ultimately help us bring professional-quality translation to underserved languages.
Finally, and perhaps most interesting to me personally, is moving beyond text to genuine multimodal understanding. Real human communication isn’t just words in isolation. It’s gesture, intonation, context, and physical presence. The real future of AI translation lies in processing true multimodal context and striving for feeling and emotional understanding.
Technology Evolution
When discussing how Translated’s technology has evolved since the company launched, Trombetti described:
We’ve ridden every major wave in AI over 25 years: rule-based machine translation in the early 2000s, statistical machine translation around 2004, adaptive neural machine translation with ModernMT in 2012, and, in 2024, Lara, our own large language model built specifically for translation. Lara is different because it explains its reasoning and asks clarifying questions when it encounters ambiguity.
The major shift has been from sentence-by-sentence translation to full-document understanding. Lara knows what’s been said earlier, how terminology has been used, and what the overall tone is. That contextual awareness is what gets it to human-level quality.
We also developed Trust Attention, a method that weights training data based on reliability. Not all translation data is equal. A professionally reviewed translation from a domain expert carries more signal than a crowdsourced amateur attempt. Training on larger but lower-quality datasets often produces worse results. Trust Attention lets us be selective at scale.
Major Company Milestones
When asked about Translated’s most significant milestones, Trombetti said:
Launching Lara in November 2024 was our biggest technical milestone. Professional translators consistently rated it higher than Google Translate and DeepL in blind evaluations. We followed with Lara V2 in mid-2025, improving quality by up to 46% across 50 languages. Then, in November 2025, we introduced Lara Think, a reasoning model that achieves 40% better human evaluation grades than our baseline. Most recently, in July 2026, we announced Lara V3, trained with a new technique we call ‘Learn by Doing’ that lets the model learn from its own work; it scored highest on quality of any system we tested while translating 23 times as many characters per second as Claude Fable 5.
In 2024, we also launched DVPS, the €29 million European Union project on multimodal AI. It is one of the largest Horizon Europe projects to date, bringing together 20 research institutions. The goal is to build AI systems that learn from direct interaction with the physical world by combining language, vision, and sensor data. It’s a five-year project aimed at translation challenges that current systems can’t handle.
The Vatican partnership, announced in early 2026, represents a different kind of milestone for Translated. It is one of the first large-scale applications of AI interpretation in a sacred space, serving millions of pilgrims in 60 languages. Beyond the technical achievement, it shows AI’s potential for inclusion in deeply human contexts.
Competitive Differentiation
When asked what differentiates Translated from its competitors, Trombetti emphasized:
Most translation companies buy their AI from Google, Microsoft, or OpenAI. We build ours. That gives us control over the entire stack. We can optimize for specific use cases, handle proprietary data securely, and iterate faster. It also means we compete on quality and specialization, not just price.
The other differentiator is our commitment to research. Research feeds directly into production, and production data informs research priorities. DVPS is a good example. It’s frontier research on multimodal AI, but it’s driven by real translation problems: How do you translate accurately in a noisy room with multiple speakers? How do you handle gesture and intonation?
We also pursue partnerships other companies don’t. The Vatican collaboration required understanding the Church’s mission, the acoustics of St. Peter’s, the sacred nature of the content, and the needs of millions of pilgrims. That’s work you can only do through close collaboration. We don’t just provide translation services. We work with institutions on problems that matter.
And we operate on a different time scale. While many companies optimize for quarterly results, we’re willing to spend 15 years building ModernMT or five years on DVPS. That’s partly possible because we’re founder-led and largely self-funded. We don’t answer to short-term investor pressures. We answer to whether we’re making progress toward universal understanding.
Future Success
When discussing what future success looks like for Translated, Trombetti concluded:
Success is reaching language singularity across all major language pairs by 2027 or 2028: AI-assisted translation indistinguishable in quality from top human translators working alone. We’re close, but close isn’t good enough. We need to be there consistently, across domains and in real-world conditions.
Long-term, success is a world where language barriers don’t determine what information you can access, who you can collaborate with, or what opportunities are available. A researcher in Brazil reading the latest papers from Japan. A Kenyan startup pitching to European investors without translation costs eating its budget. A pilgrim at St. Peter’s following Mass in their own language.
Ultimately, success means we’ve helped accelerate progress on other challenges because we’ve removed language as a limiting factor. We want to allow everyone to understand each other in their own languages so humanity can solve the most significant challenges, like landing on Mars or addressing climate change. Translation isn’t the end goal, but the enabler for everything else humanity needs to do together.
