English is the outlier. The models are finally catching up

An Arabic agent, built as an agent

Jais 3 introduces a suite of open-weight, Arabic-centric agentic models, and the largest of them, at 375B parameters, is the largest open Arabic-centric language model to date. The models support multi-step planning, tool use, terminal operation, and file manipulation across PDFs, spreadsheets, and presentations, which means they are built for agent work rather than chat alone. They use a 512K-token context window, excluding the 0.9B model, and the training pipeline includes continual pretraining, supervised fine-tuning, FTPO, and RLVR to preserve reasoning and English capabilities while specializing in Arabic. m:…5945

The design choice that matters is the last clause: preserving reasoning and English capabilities while specializing in Arabic. Most multilingual models are English models with other languages added, and the specialization tends to cost capability elsewhere. Jais 3 is a preprint release, so its benchmark claims are not peer-reviewed, and teams building on it should test the agent claims themselves before relying on them. m:…5945

Transcription goes wide

Otter.ai announced the addition of 27 new languages to its platform, bringing its total support to 31 languages once all planned languages are live. Thirteen of the new languages are available for use today, and 14 additional languages are scheduled to be released soon. The new multilingual transcription feature automatically transcribes conversations as speakers switch between languages, and Otter's Work Mode features, including summaries and CRM updates, function across all supported languages. m:…5886

Otter says it achieved 96% accuracy in English benchmark testing on the Open ASR Leaderboard, and that it develops and runs its speech models on its own infrastructure rather than using third-party AI services, so it does not send user audio to third parties for processing. The new language capabilities are available at no additional cost across all Otter plans. Otter has over 40 million users, and more than 86% of Fortune 500 companies use Otter, which makes a 27-language expansion a large real-world test of multilingual speech in meetings, not a demo. m:…5886

Why English-first breaks

Invisible Technologies explains why expansions like Otter's are hard to get right. Multilingual LLM training often breaks when built around English, because English is a linguistic outlier compared to most of the world's languages: it lacks grammatical gender, honorifics, and tonal features. English-built quality rubrics can leave out features like honorifics, causing rater disagreement, and generic hiring for languages like Arabic leads to dialectal mismatches and unstable model performance. Translated training data often flattens idioms and carries over English metaphors. m:…5373

The company's answer is to treat each language as its own project. It sources data at the dialect level for multilingual training, builds custom rubrics for each language before rating begins, and says model behavior regarding formality and honorifics should be set explicitly per language. Its Language Atlas covers 140+ languages and 1,700+ dialects, scores languages on 11 linguistic traits, and uses data from WALS, Glottolog, Ethnologue, and Grambank. The Atlas is a map of the problem: if your training data does not know a language's traits, your model will not learn them. m:…5373

The measured gap

A preprint benchmark for Bengali, BDBBQ, puts numbers on the gap. All evaluated models exhibit lower accuracy in Bengali, with Mistral-Nemo showing the smallest cross-lingual accuracy gap at 5.12 percentage points. Disambiguated contexts yield substantially greater accuracy improvements in Bengali, at 36.59 percentage points, than in English, at 17.71 percentage points, which suggests the models are guessing more in Bengali and benefit more from being given the context. Stereotype-aligned bias exhibits language and model-dependent variation, increasing in Bengali for Gemma 2 9B but decreasing for Mistral-Nemo, and higher model accuracy does not necessarily correspond to lower stereotype-aligned bias. m:…5947

The practical point for builders is that a model that is accurate in English can be both less accurate and more biased in another language, and the two failures do not move together. Shipping a multilingual feature on the strength of English benchmarks is a decision made without evidence about the languages being shipped. The week's releases show the industry starting to do the work properly: per-language models, per-language data, and per-language benchmarks. m:…5947 m:…5373