The 10 most important stories
Tencent released Hy4 preview, a new mixture-of-experts model for software engineering, office analysis, game development and scientific research. The model card lists 770 billion total parameters with 49 billion active per token, 78 layers and a one-million-token context; full and FP8 weights are on Hugging Face and other repositories with vLLM and SGLang deployment guidance. Sparse activation cuts the compute needed per token versus a dense 770B model, and open weights widen deployment choices, but the total resource footprint remains very large and the benchmark lead is not independently confirmed.
Judge Rita F. Lin of the Northern District of California ruled that the government's designation of Anthropic as a national-security supply-chain risk was unlawful. The 59-page order found First Amendment retaliation, lack of due process and arbitrary-and-capricious agency action, and ordered the challenged actions rescinded. The ruling creates an important limit on using national-security procurement powers against domestic AI suppliers; separate cases and a possible appeal remain.
Neutral-atom quantum computing company Pasqal completed its combination with Bleichroeder Acquisition Corp. II and began Nasdaq trading under the ticker PSQL, with about US$360 million in cash at closing. Reports recorded a first-day share rise. Public capital gives a European quantum firm room to scale, but error-correction, commercial-adoption and revenue-quality risks persist.
Ukraine's Ministry of Defence said an AI assistant answering product-support questions inside the DELTA interface is available to all users, handling about a quarter of web-widget queries, cutting support requests through that channel by about 27 per cent and reducing response wait times thirtyfold. This is a production support assistant, not evidence of autonomous combat decision-making, and the metrics are wartime government claims without independent technical validation.
After a meeting led by Vietnam, Qualcomm CEO Cristiano Amon said he plans to build Vietnam into the company's third-largest AI R&D centre worldwide. Vietnam seeks deeper participation in AI, semiconductors and advanced technology, with separate discussions recorded with Qualcomm and Samsung. With no investment value, hiring figures or milestones released, this should be read as strategic intent.
Anthropic opened a research preview of the Model Hardware Standard (MHS) to selected labs and manufacturers. The model-agnostic specification connects agents to programmable devices such as microscopes, liquid handlers and robotic arms, and can use MCP as the access protocol. Safety evaluation and best practices will be developed before an open-source release. MHS could become the interoperability layer between agent orchestration and physical equipment; it is not a certification or production-safety standard, and deterministic interlocks and human-authorised boundaries remain essential.
Germany's interior minister said eight NATO countries on the Baltic Sea border plan to form a task force for rapid information sharing and coordination on drone and hybrid threats. Reports linked infrastructure incidents and recent drone discoveries, though attribution for specific incidents remains disputed or unconfirmed. This is an organisational plan, not a deployed AI or counter-drone system.
More than 100 technology, finance and infrastructure companies called in an open letter for cyber defence to become a leadership priority, for faster trusted access to advanced defensive models, and for better coordination, warning that AI-assisted attacks are becoming widespread and sophisticated. Enterprises should prioritise identity hardening, least privilege, fast patch validation and incident exercises for agent-enabled systems. It is a policy recommendation, not a binding control standard.
About 80 supporters, including actors and voice performers, backed the Save Our Voices Now campaign in the UK, demanding legal rights and consent protections against unauthorised AI voice cloning. The UK Government has previously found that current law may not cover all harmful digital replicas. This is an advocacy campaign, not legislation, and its figures on cloning speed and scam exposure are not independently verified.
