Summary
This weekly AI news roundup from the AI Search channel covers a dense slate of open-source releases and research advances. Nvidia’s RD model enables real-time 3D human motion generation from text prompts with infinite streaming, released with code at roughly 765 MB per model and compatible with the Unitree G1 humanoid robot. MobileOne brings Alibaba’s video generation model to smartphones, producing 5-second 480p clips in about 20 seconds through aggressive pruning and a three-step generation process.
On the model compression front, Bonsai 27B shrinks Qwen 3.6’s 27-billion-parameter weights to just 3.9 GB using ternary quantization — small enough to run on an iPhone 17 Pro — down from the original 56 GB. OpenAI is noted as disclosing a new model called GPT Red, designed to probe and attack other AI systems. Google releases an open-source model for precise head and face control, and a separate project converts full songs to per-instrument MIDI. Nvidia also ships an update to its PID upscaler, now at version 1.5 with improved detail and color fidelity.
Additional items include humanoid robot demos, an MCP integration for Higgsfield enabling generative media workflows from within tools like Claude Code, and the Kimi K3 model release. The episode provides a fast orientation to what launched this week across open-source AI, on-device inference, and multimodal generation.
📺 Source: AI Search · Published July 19, 2026
🏷️ Format: Roundup
