AI Update: Open Models, Open Doors

Chinese President Xi Jinping elevated artificial intelligence further in Beijing’s strategic agenda with a keynote at the World AI Conference in Shanghai. Xi called for “open-source, openness, collaboration and sharing”, alongside human control of AI, coordinated international standards and capacity building for the Global South. He also opposed expanding national-security restrictions in ways that put one country’s security above others, and called for coordination through the United Nations on AI strategies, governance rules and technical standards.

Moonshot AI has introduced Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with native vision and a one-million-token context window. Only 16 of its 896 experts activate for a given task, an architecture Moonshot says improves scaling efficiency by about 2.5 times over Kimi K2. The Chinese company acknowledges that K3 still trails the strongest proprietary systems overall, but reports competitive results on coding and knowledge-work tests. K3 is already available through Kimi’s products and API, with full model weights scheduled for release on July 27.

Mira Murati’s Thinking Machines Lab has released Inkling, its first open-weights model. The multimodal mixture-of-experts system has 975 billion total parameters, activates 41 billion at a time, supports a one-million-token context window and was trained on 45 trillion tokens spanning text, images, audio and video. Thinking Machines says Inkling is designed as a customizable foundation rather than the strongest model overall, and has made it available for fine-tuning through its Tinker platform. A smaller version with 12 billion active parameters is also in preview.

London-based robot maker Humanoid has become a unicorn after raising $150 million in the first tranche of a Series A round that values it at $1.2 billion before the new capital. The company is seeking another $80 million to $100 million by September. The financing backs an unusually aggressive deployment plan: Humanoid has a binding agreement with Schaeffler targeting a four-digit number of wheeled robots across the manufacturer’s facilities by 2032, with the first systems due in German factories late this year.

xAI is testing an unusual legal response to model misuse, suing South Carolina resident Terry Harwood over allegations that he used Grok to create sexualized deepfakes of adults and minors. The company says Harwood violated Grok’s terms of service by uploading ordinary images and trying to turn them into explicit material. xAI is seeking damages and a permanent order barring him from Grok. Its complaint says xAI suspended 52,222 accounts and sent 73,604 reports to the National Center for Missing & Exploited Children in 2026, contributing to at least 244 arrests.

Researchers at the University of Tokyo and Max Planck Institute for Informatics have developed ExposeAnyone, a personalized deepfake detector that looks for unnatural relationships between speech and facial movement instead of pixel artifacts. After pretraining on more than 450 hours of video, the system can adapt to a person using about 60 seconds of authentic footage. It averaged more than 95 percent accuracy across established datasets and detected almost 95 percent of manipulated videos in a harder Sora 2 test, though its computational requirements currently rule out real-time use.