multilingual

NVIDIA released NemotronLabs RE-USE, a 9.6M-parameter multilingual universal speech enhancement model that restores speech quality across diverse degradations and sampling rates, outperforming larger open-source USE models while being ready for commercial use.

The paper introduces SERIAL-7, a Korean–English benchmark with an executable verification framework that evaluates LLMs on deriving, realizing, and preserving machine-verifiable outputs across seven representations and 43 atomic constraints, revealing significant gaps in structured generation reliability across ten models.

Nedimo presents an evidence-grounded, multilingual pipeline for generating low-arousal AI sleep stories, separating deterministic planning from LLM generation with attention tapering, validation guardrails, and TTS loudness normalization, while positioning the system as a relaxation tool without clinical efficacy claims.

Meta Superintelligence Labs researchers formalize 'adaptive personalization under epistemic uncertainty' and introduce a multilingual benchmark revealing a recognition–generation gap in LLMs, where models default to exploit-first behavior despite appropriate strategy identification; targeted training on internal strategy representations narrows the generalization gap.

Datalab 发布 Surya OCR 2.5，一个 689M 参数的 OCR 模型，在 OmniParseBench 上取得 85.9 分（1B 以下开源模型最高），支持 91 种语言、布局分析与表格转 HTML，吞吐量在单张 H100 上达每秒 4.5 页。

Mistral 发布 Mistral Large 4（Le Chonk）：一个 1 万亿参数、49B 激活参数的原生多模态 Mixture-of-Experts 模型，具备 1M 上下文窗口、可配置推理力度、160+ 语言支持，并以 Mistral Ecosystem Open-Weight License 开源权重发布。

ElevenLabs 推出 Cue，一款强制对齐模型，可将音频中的每个字符精确到亚毫秒级时间线，准确度超过其测试的所有对齐器，覆盖 29 种语言，且一小时内音频的对齐耗时不到一秒。

Beamfy, an AI vision assistant for blind people supporting 74 live languages, has launched on Product Hunt.

NVIDIA released Nemotron 3.5 Transcribe 30B-A4B, an open-weight offline speech and audio transcription model supporting 40 languages with a hybrid MoE architecture that activates ~4.3B parameters and handles recordings up to one hour.

Nvidia 发布了免费开源权重语音转写模型 Nemotron 3.5 Transcribe（30B 参数，支持 40 种语言）以及 Granary 开源语音数据集 v2，瞄准会议记录和医疗转写等企业应用场景。该模型声称在 Open ASR 等基准上超越竞争对手，且可在单芯片上运行。

NVIDIA released Nemotron 3.5 Transcribe, an open-weight promptable model for multilingual transcription, speaker diarization, translation, and audio Q&A with top leaderboard results, alongside the Granary v2 data pipeline and NeMo fine-tuning recipes for domain adaptation.

Muse is now available in Singapore, marking an international expansion for the device, which supports the country's four official languages.

ETC-Bench is the first comprehensive multilingual benchmark for evaluation-side transcript canonicalization, covering 181 phenomenon types across 12 languages, along with a new SENA metric and a taxonomy-guided agent method (TACS) that outperforms baseline systems.

Researchers train ModernBERT-based Icelandic encoders from 12M to 1B parameters on 9B tokens of open data, showing that a 1B model trained from scratch beats IceBERT on all eleven tasks, and that warm-starting from an English Ettin encoder yields the best scores across every benchmark while highlighting the importance of vocabulary transplantation.

Liquid AI released LFM3-Audio-2B, a compact 2B-parameter audio model that runs locally on laptops and edge devices, supporting voice chat, multilingual speech across 18 languages, audio understanding, and tool calling with an in-house audio encoder and LiquidCodec tokenizer.

ARMOR-DP is a modular framework for multilingual BabyLM that dynamically adjusts language sampling based on per-language validation loss, showing that the weakest language shifts during training and that dynamic protection outperforms fixed Chinese-protection baselines.

该论文提出 SEDATE（选择性检测与定向编辑），一种基于检测-编辑-验证策略的免训练方法，通过定位跨语言共享的“攻击性”神经元来提升大模型的安全性，并在7种语言和4个模型家族上验证了其优于基线的效果。

AdaptTok-Indic 提出一种面向印地语系语言的脚本感知自适应分词方法，结合 Unicode 归一化预处理、显式印度语连字处理和基于脚本路由的字节级边界预测，在孟加拉语和印地语上取得比 SentencePiece BPE 更低的 token 生长率和更好的下游表现。

throxy (YC P25) 推出 AI 原生销售服务，为制造业、物流、建筑等 B2B 公司预订合格销售会议，新增按客户语言（如法语、德语）转接对应语言 SDR 通话的功能，按成功发生的会议付费。

This paper causally analyzes how multilingual LLMs represent negation across English, French, Vietnamese, and Chinese, showing that a shared, causally controllable negation direction—and even a single sparse autoencoder feature per model—can induce or remove negation across languages and generalize to unseen ones.