September 2026 newsletter

News, jobs, community questions and more about machine translation from September 2026

🗞️ News

GPT-6 Astra leads machine translation benchmarks, but remains below human-level translation quality, according to tests by RWS and the Last Translation Benchmark.
The newly implemented EU AI Act does not require translation tools to disclose AI use, unlike text generators and rewriters.
HarperCollins’s new manga line, KAZE, will not use AI translation, in support of human translators.
Anthropic adds SynthID-Text watermarking to Claude’s text output, which is kept when the text is copied or reformatted, but not when it’s translated or rephrased.
Google limits its AI sign language translation tool to casual conversation, citing accuracy and liability concerns.

More news →

🚀 Launches

Cohere launches North Small Translate, an open machine translation model with 218B parameters (25B active) and a 16k context length that outperforms other open MT models as well as Google Translate and DeepL across 50+ languages.
DeepL Voice launches real-time speaker identity and tone preservation for live meetings, keeping individual vocal characteristics and emotional delivery across 12 languages.
Google Translate launches iOS and Android upgrades: Android users can keep translating while using other apps or with the screen locked, and iPhone users can hold the phone to their ear to listen to the translation privately.
Google launches Gemini 3.8 Live, with real-time speech-to-speech reasoning and automatic language detection across 97 languages for code-switching during live dialogue.

More launches →

🛠️ Community launches

LST — Local Subtitle Translate launches on the Chrome Web Store, using Ollama to translate Netflix subtitles locally in real time.
i18n-keyless, an open-source internationalization library with automatic translations, skips the usual setup of translation keys and per-language files.
Manga Translator detects speech bubbles, runs OCR, translates into 30+ languages, and re-renders text back into the original bubbles.
AITnovel, a paid AI translation service for entire novels, supports TXT, EPUB, and PDF, with editable glossaries for names and terminology.
ALST, an open-source Android app, translates on-screen text and overlays the translation directly on top of the original in real time, using Gemini Vision with an offline backup mode.
Champollion, an open evaluation and development suite for low-resource languages, aims to support community projects like Masakhane and AmericasNLP.
cinemacc-subtitle-skill, an open-source add-on for AI assistants like Claude, builds a glossary of names and terms before translating subtitles, to keep them consistent across a file.
An open-source video dubbing pipeline combines Whisper transcription, translation, and VoxCPM2 voice cloning, demoed on an English-to-Turkish dub.
ConveyThis open-sources a GitHub repo for translating multilingual websites and a Claude skill that translates static sites while preserving HTML structure, SEO metadata, hreflang, canonicals, and sitemaps.
An AI news app translates live news from 150+ countries, summarizes 2,000-word articles into 3 bullets, and compares coverage across 60,000 publishers.

More community launches →

📚 Research

A comparison between a general-purpose model, Qwen3-8B, and two dedicated translation models, HY-MT1.5-1.8B and TranslateGemma-4B, on a budget GPU finds the dedicated models faster, but less reliable at preserving numbers and code.
The Last Translation Benchmark paper collects 3,456 hard-to-translate examples that break state-of-the-art models, arguing machine translation is far from solved.
A year translating novels with local AI models yields lessons on keeping glossaries, splitting the work across two models, and why invented words still break machine translation.
A developer mixing two languages in the same sentence finds that hinting Whisper’s API with three candidate languages, not the two actually spoken, stops it from dropping words.
AMTA publishes a framework for evaluating translation QE systems.
RWS TrainAI launches M-GATE, a benchmark of 70 AI models across 30 languages, finding top reasoning models often score worse than a random guess on non-English grammar.
DATAmundi tests AIDA multi-agent frameworks on terminology-constrained translation, evaluating strict glossary adherence for enterprise MT.

More research →

❔ Questions

Why is Reddit’s translation feature such a mess? Can’t they plug in other translation APIs?
Are Meta’s OMT-NLLB (Omnilingual MT) model weights publicly released anywhere?
How should I position myself for a job search (non-tech background, 20+ years of cross-lingual work, two years diagnosing LLM translation failures)?
Which small local LLMs work best for translation?
What’s the best AI tool for translating dense philosophy and psychoanalysis books?
How do I translate a DRM-protected EPUB from English to Italian?
Is Google Translate’s voice recorded word by word, or broken down into phonetic sounds?
Is live translation during meetings actually useful? (OLVA team asking for feedback)
Is there a multilingual voice transcriber for WhatsApp?
How do you handle documents that need re-translating every time they’re updated?
DeepL’s quality has dropped. Is Google Translate better now?
Which rule-based machine translation system supports Portuguese to English?