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

Google released Gemma 4, their latest family of open-source AI models, which can run on consumer hardware.

Gemma 4 is fully open under the Apache 2 license, which has very minimal restrictions.

Gemma 4 comes in four sizes: E2B (2 billion parameters), E4B (4 billion parameters), 24 billion parameters (mixture of experts), and 31 billion parameters (dense model).

The E2B and E4B Gemma 4 models are suitable for running on phones, Raspberry Pi, or other edge devices.

The 24 billion parameter Gemma 4 model is a mixture of experts model, where only 3.8 billion parameters are active when in use, making it efficient.

Gemma 4 is natively trained on over 140 languages.

Generative World Renderer is an AI that pulls GBuffers (graphics data) from real AAA video games to allow editing and restyling gameplay with text prompts.

The GenSearcher framework is model agnostic and can work with any image model, outperforming base models in accuracy and visual correctness.

TokenDial allows users to control specific elements like emotion, style, or motion in video generation using slider controls.

Hybrid Memory for dynamic video world models is a framework that helps world models remember objects in a scene, maintaining consistency even when the camera pans away and back.

The HM World dataset, used to train Hybrid Memory, contains around 59,000 high-quality video clips with subjects moving in and out of view.

DreamLite can generate a 1024x1024 image in approximately 3 seconds on an iPhone 17 Pro.

Claude Code Code leak provides insights into how to program a coding agent to be more efficient and powerful, as the competitive edge in coding assistants lies in clever engineering around the underlying model.

The quality of Wan 2.7 is still behind ByteDance's Seedance 2.0 and Cling 3.