This week in AI has been packed with exciting releases and advancements across multiple domains. Anthropic launched Claude Opus 4.7, a model excelling in complex software engineering and autonomous workflows, though it shows mixed performance in some areas compared to its predecessor. Alibaba introduced Quen 3.6, a 35-billion-parameter open-source mixture of experts model that outperforms peers in coding and reasoning tasks. Additionally, Alibaba’s ATH Lab unveiled Happy Oyster, an open-ended 3D world generator similar to Google’s Genie 3, enabling real-time interactive environments with versatile character and scene generation.
In the realm of 3D modeling and video generation, several breakthroughs emerged. Annigen can create fully rigged 3D models ready for animation from a single image, outperforming competitors in skeleton estimation. Motif Video 2B is a lightweight, efficient video generator requiring significantly less data and compute while maintaining high-quality outputs. Nvidia released LRA 2, a tool that converts videos into consistent 3D point clouds for robot training simulations, and Tencent announced HY World 2.0, a multimodal model generating interactive 3D worlds from text, images, or videos, with open-source components forthcoming.
Efficiency and accessibility improvements were highlighted by the release of Turnary Bonsai, a family of ultra-efficient 1.58-bit language models that are dramatically smaller yet competitive with larger models, capable of running on consumer devices and mobile chips. OpenAI introduced GPT Rosalind, a specialized reasoning model designed to accelerate life sciences research by integrating literature review, experimental planning, and data analysis, connecting to over 50 scientific databases to streamline drug discovery and genomics workflows.
On the robotics front, Unree showcased their H1 humanoid robot sprinting at record speeds, while the second robot marathon in Beijing demonstrated significant advancements in autonomous humanoid running. Leju Robotics revealed the first automated production line for humanoid robots, capable of assembling one unit every 30 minutes with high automation and quality control, signaling a shift toward industrial-scale robot manufacturing. Adobe contributed with Token Relight, an AI tool offering precise, continuous control over lighting attributes in images, enhancing creative workflows.
Finally, Google released Gemini 3.1 Flash Text-to-Speech, a highly expressive TTS model supporting over 70 languages and capable of nuanced emotional and rhythmic control via metatags, surpassing competitors like 11 Labs. Bite Dance introduced Omni Show, an AI for generating realistic user-generated content videos with accurate person and product representation, customizable audio, and pose control, aimed at marketing and influencer content creation. The week also saw the launch of Game World, a benchmark for evaluating AI performance in browser-based games, highlighting ongoing challenges and progress in agentic AI gameplay.
