CAS unveils SpikingBrain1.0, a brain-inspired LLM trained on MetaX GPUs with hybrid-linear attention, reporting 100× TTFT at 4M tokens and high sparsity for low power.
Built by EPFL, ETH Zurich, and CSCS, Apertus is a fully open LLM trained on 15T tokens across 1,000+ languages, with EU-aligned transparency and access via Swisscom and Public AI.
Microsoft introduces MAI-Voice-1 and MAI-1-preview, a mixture-of-experts model trained on ~15,000 Nvidia H100s, with GB200 compute active and targeted Copilot rollout.
Google Translate broadens into learning with AI-personalized practice and Gemini-based live conversations, backed by 1T monthly translated words across Google services.
NotebookLM’ Video Overviews expand to 80+ languages, offering native-language video and audio overviews with cultural adaptation and Gemini-powered accuracy.
xAI open-sources Grok 2.5 and pledges Grok 3’s release in six months, highlighting transparency, enterprise benefits, and global AI rivalry.
OpenAI CEO Sam Altman addresses GPT-5’s launch issues, such as the misleading chart and real-time router failure, and outlines upcoming improvements for user experience.
Google’s Jules AI coding agent exits beta with Gemini 2.5 Pro, asynchronous execution, audio changelogs, structured pricing, GitHub integration, and global adoption.
Anthropic’s latest release, Claude Opus 4.1, improves software engineering accuracy to 74.5%, enhancing research and data analysis tasks for businesses and developers.
