Descriptions:
AI Search’s weekly roundup covers an unusually dense slate of announcements. The headline items are Anthropic’s release of Claude Opus 4.7 and Alibaba’s open-source Qwen 3.6 model drop. Alibaba’s ATTH lab also unveils Happy Oyster, a real-time interactive 3D world generator positioned as a competitor to Google’s Genie 3, while Tencent announces HY World 2.0, a multimodal world model that generates explorable 3D environments from text, images, or video input. Nvidia contributes LRA 2, which converts ordinary video footage into interactive 3D Gaussian splat environments suitable for robot training in Isaac Sim.
On the efficiency and open-source front, the video profiles Ternary Bonsai — a new family of 1.58-bit language models at 1.7B, 4B, and 8B parameters that are nine times smaller than standard 16-bit equivalents. The 8B variant is just 1.7 GB yet achieves benchmark performance close to Llama 3.1 8B and Qwen 3 8B while exceeding 100 tokens per second on consumer hardware, including mobile chips. Additional coverage includes a Prompt Relay technique for seamless multi-scene video generation on top of Alibaba’s Wan, Google’s new emotion-aware text-to-speech system with meta-tag support, OpenAI’s new scientific research AI, several humanoid robot demonstrations, and a new tiny open-source video model.
The video is a solid weekly digest for AI professionals tracking competitive model releases, open-source efficiency research, and the emerging race to build real-time 3D world simulation capabilities across Alibaba, Tencent, Google, and Nvidia.
📺 Source: AI Search · Published April 19, 2026
🏷️ Format: News Analysis
