Summary
AI Search delivers a wide-ranging weekly roundup covering open-source releases across video generation, 3D reconstruction, image editing, voice, and robotics. Leading the coverage is Evoke, a 14-billion-parameter open-source interactive world model that generates video in near real-time from image inputs and joystick controls in just three diffusion steps, with suggested applications in synthetic robot training data for scenarios like hospital navigation or underwater exploration. The model and all training-stage checkpoints are released under Apache 2.0.
Also featured is 4D Anyone, which reconstructs moving 3D Gaussian splat characters from a single input video and is compared favorably against ReCapMaster and Trajectory Crafter. SenseNova U 1.58B arrives as a new open-source image generator capable of photorealistic outputs and poster design. DeepSeek drops a new model with vision capabilities, and Qwen VideoEdit — Alibaba’s approach of running QwenImageEdit frame-by-frame through the Wand video pipeline — enables text-driven video editing from an existing video. A compact text-to-speech model designed for low-end hardware is also covered, alongside Geo Weaver, a long-video 3D scene reconstruction system that stitches overlapping chunks into globally consistent point clouds.
The robotics segment highlights a robot breaking the human world record for highest jump and fastest sprint, plus a transformer robot and a tennis-playing robot. The episode functions as a useful weekly digest for practitioners tracking open-source releases across modalities.
📺 Source: AI Search · Published August 23, 2026
🏷️ Format: Roundup
