The agentic era finally learns to speak more than English

An Arabic agent suite, open weights

Jais 3, a preprint introducing a comprehensive suite of open-weight, Arabic-centric agentic models, ranges from 0.9B to 375B parameters and is designed for complex tasks like tool use, terminal operation and document manipulation. The 375B parameter model is the largest open Arabic-centric language model to date, the authors say, and the models support multi-step planning, tool use, terminal operation and file manipulation across PDF, spreadsheet and presentation files. They utilize a 512K-token context window, excluding the 0.9B model, which matters for agents that must work through long documents in Arabic. m:…5945

The training pipeline includes continual pretraining, supervised fine-tuning, FTPO and RLVR to preserve reasoning and English capabilities while specializing in Arabic, the authors say, and the release provides full transparency through released weights, training recipes and audited benchmarks. That last point is what makes Jais 3 different from a translated model: it is built to reason and act in Arabic, with English preserved rather than dominant, and the training recipe is public so others can reproduce or extend it. For developers building localized AI agents in the Arab world, the suite offers something that did not exist before at this scale. m:…5945

Meetings in 31 languages

Otter.ai has launched 27 new languages, bringing its total support to 31, and introduced automatic multilingual transcription that transcribes conversations as speakers switch between languages. Thirteen of the new languages are available for use today, the company says, and 14 additional languages are scheduled to be released soon. The platform processes audio on its own infrastructure, the company says, and does not send user audio to third parties for processing. Its 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 reports over 40 million users, with more than 86% of Fortune 500 companies using the platform. The new language capabilities are available at no additional cost across all plans, the company says. The expansion aims to address collaboration challenges for international knowledge workers, 42% of whom cite language barriers as a top hurdle, according to the company's summary. The business logic is straightforward: the meeting product that handles the languages a global team actually speaks wins the global team. m:…5886

Teaching breaks first in low-resource languages

A preprint on multilingual reasoning distillation identifies a failure mode that explains why low-resource languages lag. Teacher drift occurs when a teacher model follows a student's incorrect reasoning steps, and it tends to occur earlier in lower-resource languages, the authors say. Their method, TOPSD, determines where and whether teacher supervision should be applied to each reasoning trajectory, using teacher-student likelihood to estimate where drift begins, and applies distillation only to prefixes that pass verification. The method consistently outperforms conventional on-policy self-distillation across ten languages and three model scales, with the largest gains in low- and medium-resource languages. m:…5411

The mechanism is simple to state and hard to execute: verify whether the teacher can recover the correct answer from a selected prefix, and supervise only the prefixes that pass. On-policy self-distillation improves reasoning by training a student on its own reasoning using a teacher model, but if the teacher follows the student down a wrong path, the student learns the wrong path more confidently. TOPSD's contribution is knowing when to teach, which is the right question for multilingual work, where the teacher's own competence varies by language and the student has the least margin for error. m:…5411

Voice cloning goes small and local

Zyphra Research released ZONOS3-Mini, a compact zero-shot voice-cloning text-to-speech model with a sparse Mixture of Experts architecture, 100 million active parameters and 800 million total parameters, licensed under Apache 2.0. It runs in real time with less than 570 ms of time to first audio in the browser on a Pixel 9, and less than 150 ms on an M3 MacBook Air, the company says. Zyphra trained it on 112 Nvidia H100s in 2.5 days, and provides runtime support via SGL, C++ and WebGPU, with servers via SGLang and vLLM. m:…6355

Zyphra developed its own audio codec to improve speech generation fidelity and reduce word error rates, and says the model ranks first among open models on its own Zyphra TTS Evaluation v2 benchmark, a claim buyers should treat as company-reported. Google's Language Explorer, built on the open-source LinguaMeta dataset as part of the 1,000 Languages Initiative, shows the scale of the remaining gap: Google Translate started in 2006 with two languages, now supports nearly 250 languages and processes about a trillion words translated every month. The tools are arriving, but the long tail of languages still depends on data collection efforts like Project Vaani in India and Project Waxal in Africa. m:…6355 m:…3371