Thread — Secure, context-aware translation for office use

The problem

The language barrier is everywhere in Europe.

Wherever local services meet people who don’t yet speak the language, the same gap appears — and generic translation can’t handle the institute-specific terms that decide outcomes.

🏛️

Governmental institutes

Ausländerbehörde, Jugendamt and other offices issuing binding decisions.

🤝

Social companies

Organizations working day-to-day with immigrants and refugees.

🎓

Adult education

Courses for learners with a migration background.

🏥

Hospitals

Private and public care where misunderstanding carries real risk.

🧒

Kindergartens

Early-years settings across Germany communicating with parents.

🚚

Logistics companies

Fleets where up to 80% of drivers come from other countries.

⚖️

Advisors

Tax and legal professionals translating sensitive client matters.

🏢

SMEs across borders

Companies with offices in different countries — e.g. Germany and Japan.

Solution

One platform for context-aware translation, in real time.

A specialized service for seamless communication between local services and immigrants — on mobile and desktop, with institute-specific, secure translations.

✓
Context-aware, not word-for-word.
Reads the whole exchange so terms like Bescheid, Frist and Widerspruch keep their legal weight.
✓
Live chat across languages.
A self-explaining chat function — each side writes in their own language and reads in it too.
✓
Document translation.
Drop in a Bescheid or form and get an accurate, terminology-checked translation back.

💬

Real-time chat

Two-way conversation at the counter, no interpreter scheduling required.

📄

Documents

Translate official letters and forms with key terms preserved.

📱

Mobile & desktop

Use every feature from the app or the web platform.

🔒

Secure by design

Institute-specific, privacy-conscious translation on EU infrastructure.

Technology · In development

We’re building DoLLaMo — a translation model made for officialese, not the open web.

Today Thread runs on tuned engines and a domain glossary. Next, we’re training our own: a specialized large language model built from the ground up on bureaucratic, legal and administrative language, so the terms that decide an outcome survive the translation instead of getting flattened.

Trained on domain corpora. EU legislative and terminology sources — Europarl, JRC-Acquis, IATE — alongside WMT benchmarks and specialist medical text.
Terminology fidelity over fluency. Optimized for faithful domain translation and measured on BLEU against expert human translations, not generic web text.
Learns continuously, with people in the loop. Caseworker corrections feed back into the model, keeping it current with evolving Verwaltungssprache.
Open by default. Model weights, architecture and datasets released as open science — trained on EuroHPC compute, GDPR-aligned with synthetic data where real data is sensitive.

+20–30%

target BLEU gain vs
general engines

−30–50%

less human
post-editing

<500ms

translation
latency

DoLLaMoModel card · targets

Architecture

Transformer encoder–decoder, domain-specialized · multi-head attention, 6–12 layers

Objective

Faithful domain translation — cross-entropy, BLEU-evaluated

Training data

Europarl · JRC-Acquis · IATE · WMT · PubMed / clinical

Languages

German ⇄ 30+

Learning

Continuous + human-in-the-loop

Compute

EuroHPC JU GPU allocation

Access

Open weights, architecture & datasets

Privacy

GDPR-aligned · synthetic data where needed

Reviews

Trusted by the people doing the work.

4.2 ★★★★

on the Google Play Store · 68 reviews

We’re continuously collecting feedback and improving.

Partnerships

Backed by the people building the ecosystem.

Accelerator

Avant Now Accelerator

Program member

NVIDIA Inception

Cloud startup program

IONOS Cloud

Co-funded by

European Union

Supported by

Berlin Senate

Regional support

Startup Lausitz

Real-time language translation that keeps the meaning.