The AI news of the day centers on a wave of open-weight model releases and infrastructure bets: Meta's new AI research group shipped Muse Glimmer, a 30B-parameter Apache 2.0 model built for local, always-on agents, while Nvidia pushed out a smaller, faster Nemotron 3.5 Lightning model alongside a router to direct workloads to it. Mistral countered with a plan to build up to 1 GW of AI compute capacity in Europe by 2030, a sovereignty play against US cloud dependence. On the enterprise side, Gartner projects AI-optimized cloud infrastructure spending to grow 96% this year to $42 billion, with inference now outpacing training as the dominant driver, and OpenAI is reportedly maneuvering against Anthropic's Claude Cowork. Security teams also got a reminder of AI's dual-use nature: Rapid7 disclosed a chained SharePoint exploit combining a JWT validation flaw with a .NET deserialization bug, while Anthropic said it will start watermarking Claude-generated text.

• Meta ships Muse Glimmer, a 30B open-weight model for local agents
• Nvidia's Nemotron 3.5 Lightning targets faster, smaller inference
• Mistral plans up to 1 GW of European AI compute by 2030
• Gartner: AI cloud infra spend to grow 96% to $42B, inference now leads
• Rapid7 chains two CVEs into a SharePoint RCE exploit