Summary
AI Search’s weekly roundup covers one of the busiest stretches in recent AI history, anchored by the US government’s export control order pulling Anthropic’s Claude Fable 5 and Mythos 5 from global access and the cascade of Chinese open-source model releases that followed.
On the model front, Kimi K2.7 from Moonshot AI arrives as a one-trillion-parameter mixture-of-experts model (32 billion active parameters) that benchmarks close to closed-source leaders like GPT-5.5 while claiming improved reasoning efficiency and reduced overthinking. ZAI’s GLM 5.2 (744 billion parameters, 1-million-token context window, MIT license, open weights imminent) and Minimax M3 — which tops Artificial Analysis open-source leaderboards ahead of Kimi and DeepSeek V4 — also debut. Google releases a fast open-source model alongside a free real-time translation tool.
On the multimodal and robotics side, Scale 2 (also from ZAI) is highlighted as the strongest open-source tool for video motion transfer, matching closed-source Cling 3 in quality and outperforming it on multi-character scenes at 81 GB total model size. The roundup also covers OSCAR, a robot world model designed to simulate physical outcomes across diverse robot hardware, and an actionable world representation system for building controllable 3D models of deformable and articulated objects from depth video. Luma Agents rounds out the coverage as a creative workflow platform for video producers. A new top-tier text-to-speech generator also receives a mention.
📺 Source: AI Search · Published June 14, 2026
🏷️ Format: Roundup
