OpenAI's account says GPT-5.6 Sol will launch publicly this Thursday alongside Terra and Luna. Details in the candidate story are still sparse, but the timing points to a near-term model rollout that teams should watch for product, API and evaluation updates.
Tencent released full weights for Hy3, a 295-billion-parameter mixture-of-experts model aimed at STEM reasoning. The model is published under Apache 2.0, making it easier for researchers and engineering teams to test, adapt and compare against closed frontier systems.
Seasonic's PSU calculator listed unreleased Nvidia RTX 50 Super-series graphics cards with higher total graphics power than current models. The entries suggest the refresh could raise board power by roughly 10 to 17 percent, which matters for workstation builds, PSU sizing and thermal planning.
Lahore Traffic Police introduced drones using AI to monitor traffic flow, congestion and road incidents in real time. The system is meant to give traffic authorities live situational data, making this a local smart-city deployment rather than a general policy story.
CISA added four exploited vulnerabilities to its Known Exploited Vulnerabilities catalog, including Adobe ColdFusion, Joomla and Langflow issues. The Langflow item is especially relevant to AI-agent builders because exposed visual workflow systems can become high-value entry points.
Researchers disclosed GhostLock, a Linux kernel flaw described as present across mainstream distributions for years. The issue can allow local privilege escalation to root and container escape, so platform teams should track vendor patches and prioritize shared hosts and container infrastructure.
Figma acquired the team behind a Y Combinator-backed vibe-coding product that later shifted toward agent creation. The move shows design platforms continuing to absorb AI-native app-building workflows rather than leaving them as separate developer tools.
French AI startup ZML released ZML/LLMD, software intended to make inference faster across a range of AI chips. For engineering teams, the useful angle is portability: lower-cost inference depends on software layers that can target more than one accelerator stack.