Descriptions:
Kat Kampf and Ammaar Reshi from Google DeepMind bring two simultaneous model announcements to the AI Engineer conference: Gemini 3 Pro, described as Google’s most capable and intelligent model to date, and a new image generation model featuring substantially improved text rendering, visual consistency, and multilingual localization.
Gemini 3 Pro is positioned around two core strengths: strong UI and aesthetic design understanding — generating polished websites with appropriate typography, animations, and layout in a single prompt — and significantly improved agentic tool-calling for complex coding tasks. SWE benchmark results shown on stage place Gemini 3 Pro substantially above previous Gemini versions and competing models in agentic scenarios. The new image model introduces consistent character rendering across images with up to 14 people, accurate text rendering across dozens of languages, prompt-based focus and depth-of-field control, and a wide range of supported aspect ratios including banners and wallpapers.
The majority of the session is live demos built in Google AI Studio. Reshi demonstrates a vibe-coded comic book story generator that incorporates uploaded photos of the presenters, rendered with consistent character likeness, accurate speech-bubble text, and scene-aware narrative generation. Kampf showcases a shader-animated website produced from a minimal prompt, making a deliberate point that Gemini 3’s design sensibilities move beyond the stereotypical purple-gradient aesthetic associated with earlier AI-generated UIs. Both models are accessible immediately via the Gemini API in AI Studio.
📺 Source: AI Engineer · Published December 15, 2025
🏷️ Format: Keynote Launch
