Google’s open-source AI, Claude Code leaked, new Wan, new Qwen, image gen on phone: AI NEWS

Key Concepts

• Multimodal AI: Models capable of processing and generating multiple types of data (text, image, audio, video) natively.
• Edge AI: AI models optimized to run locally on consumer hardware (phones, Raspberry Pi) with low latency.
• Agentic Frameworks: Systems that allow AI to act as an autonomous agent, using tools, memory, and planning to complete complex tasks (e.g., coding, graphic design).
• Latent Geometry/World Models: AI architectures that understand 3D spatial relationships and physics to ensure consistency in video generation.
• Parameter Efficiency: Techniques like Mixture of Experts (MoE) and quantization (4-bit) to reduce model size while maintaining performance.

---

1. Open-Source Model Releases

• Google Gemma 4: A new family of open-source models (Apache 2 license) ranging from 2B to 31B parameters.
  • Architecture: Includes tiny models (2B/4B) for edge devices and larger models (24B MoE, 31B dense) for high performance.
  • Capabilities: Native multimodal support (text, image, audio) and massive context windows (120k–256k tokens).
• Netflix "Void": An open-source model for Video Object and Interaction Deletion. It allows users to remove objects from videos while realistically filling in the background. Requires high-end GPUs (22GB model size).
• Alibaba Qwen 3.5 Omni & 3.6 Plus:
  • 3.5 Omni: A real-time multimodal model capable of analyzing video/audio/text simultaneously.
  • 3.6 Plus: Features a 1-million token context window and enhanced agentic coding abilities.

---

2. Specialized AI Tools & Frameworks

• Generative World Renderer: Uses "GBuffers" (graphics data including depth, normals, albedo, and roughness) to allow text-based editing of AAA video game environments.
• Gen Searcher: A model-agnostic framework that integrates web search into image generation, ensuring factual accuracy for specific characters, locations, or data-heavy infographics.
• Token Dial: A video generation control framework that uses sliders to adjust specific attributes like motion intensity, age, or style, providing finer control than text prompts alone.
• LongCat Audio: A high-performance text-to-speech and voice cloning tool supporting 600+ languages. It excels at capturing tone and mannerisms.
• See-Through: An anime-focused tool that decomposes single images into layered, depth-aware components for animation and editing.
• Hybrid Memory (Hydra): A framework for video world models that uses "memory tokens" to maintain object consistency when items move in and out of the camera frame.
• Dreamlight (ByteDance): A tiny (0.39B parameter) image generator optimized for mobile devices (iPhone 17 Pro), capable of generating/editing 1024x1024 images in ~3 seconds.
• PS Designer: An agentic framework that generates layered Photoshop files from prompts by using an "asset collector" and "graphic planner."
• LGTM (Less Gaussian Texture More): A 3D reconstruction method that uses compact Gaussian representations with attached textures to achieve 4K resolution without excessive compute.
• Han X: A dataset of detailed 3D hand movements designed to train humanoid robots in dextrous manipulation.

---

3. The Claude Code Leak

• The Incident: Anthropic accidentally leaked over 500,000 lines of source code for their "Claude Code" agent via an npm package containing a source map.
• Key Discoveries:
  • Buddy: A Tamagotchi-style virtual pet that interacts with the user.
  • Chyros: An "always-on" background agent mode for persistent tasks.
  • Undercover Mode: A system instruction forcing the AI to act like a human developer, stripping internal metadata from commit messages.
• Significance: The leak highlighted that the competitive advantage in AI coding assistants lies in the "agentic engineering"—how the model manages memory, tool use, and task decomposition—rather than just the base model.

---

4. Synthesis and Conclusion

The AI landscape this week shifted heavily toward efficiency and agency. The release of Google’s Gemma 4 and ByteDance’s Dreamlight demonstrates a clear trend toward bringing powerful, multimodal AI to edge devices. Simultaneously, the "agentic" revolution is maturing; tools like PS Designer, Claude Code, and Qwen 3.6 show that the industry is moving beyond simple chat interfaces toward systems that can plan, execute, and maintain state across complex, multi-step workflows. The focus on 3D consistency (V-GGPO, Hydra) and factual grounding (Gen Searcher) suggests that the next frontier for AI is moving from "creative generation" to "reliable, consistent, and functional production."