Cohere Lança Família de Modelos Multilíngues Open-Weight Tiny Aya
← Voltar às Notícias Enterprise AI firm Cohere launched the Tiny Aya family of open-weight multilingual models, supporting over 70 languages and designed for on‑device use. The base model contains 3.35 billion parameters and runs on everyday hardware without internet connectivity. Regional variants target African, South Asian, and Asia‑Pacific/West‑Asia/European languages. Trained on a single cluster of 64 H100 GPUs, the models are available on HuggingFace, the Cohere platform, Kaggle and Ollama, with accompanying datasets and a forthcoming technical report. Cohere also highlighted strong financial performance and a pending public‑market plan. Ler mais Lightricks introduced an AI video model that runs locally on consumer devices, a rarity in the industry. Built with Nvidia technology and showcased at CES 2026, the model can generate 20‑second clips at 50 frames per second, supports 4K resolution and native audio, and is released as an open‑weight model on HuggingFace. On‑device operation gives creators control over their data, faster generation times of 1‑2 minutes per prompt, and reduced reliance on cloud services, addressing privacy and efficiency concerns for both independent filmmakers and large studios. Ler mais Researchers from Texas A&M, the University of Texas and Purdue University have introduced the “LLM brain rot hypothesis,” suggesting that continual pre‑training on low‑quality web text can cause lasting cognitive decline in large language models. Their pre‑print paper analyzes a HuggingFace dataset of 100 million tweets, separating “junk” tweets—identified by high engagement yet short length or superficial, click‑bait content—from higher‑quality samples. Early results show a 76 percent agreement between automated classifications and graduate‑student evaluations, highlighting the potential risks of indiscriminate data ingestion for AI systems. Ler mais Switzerland has introduced an open‑source artificial intelligence model named Apertus, positioning it as an alternative to proprietary systems such as OpenAI’s ChatGPT and Anthropic’s Claude. The model’s code, training data, weights and development details are publicly available on HuggingFace. Designed to establish a new baseline for trustworthy, globally relevant AI, Apertus was trained on more than 1,800 languages and is offered in two configurations—one with eight billion parameters and another with seventy billion. The developers say the model aligns with European Union copyright rules and respects opt‑out requests from websites, avoiding any stealth‑crawling of data. Ler mais
Tags: HuggingFace
Lightricks Apresenta Modelo de Vídeo AI On-Device Impulsionado pela Nvidia
Estudo Relaciona Treinamento de Dados de Baixa Qualidade ao Desempenho Diminuído de Modelos de Linguagem Grande
Suíça Lança Modelo de IA de Código Aberto Apertus