Social media platform X recently revealed a major leak about Google's next-generation model Gemini3.5. According to a post by blogger Pankaj Kumar, an internal test version codenamed Snow Bunny demonstrated remarkable engineering capabilities, capable of building complete applications in one go.

image.png

The leaked information indicates that Gemini3.5 has made a qualitative leap in the field of programming, supporting the generation of up to 3000 lines of executable code with a single prompt. To address different use scenarios, Google seems to have prepared multiple specialized models: Fierce Falcon focuses on pure speed and logical reasoning, while Ghost Falcon excels at handling UI design, visual effects, and audio creation.

In terms of core architecture, Gemini3.5 introduced "System2 Reasoning," allowing the model to engage in deep thinking before answering. This newly added "Deep Think" mode enables it to score as high as 80% in difficult logic tests, far surpassing its competitors. Leaked data even claims that Gemini3.5's performance has exceeded GPT-5.2 and Claude Opus4.5.

Although these details have not yet been officially confirmed by Google, code snippets in the leaked information already show variables related to gemini-for-google-3.5, indicating that the model's API is almost ready for launch.

Key Points:

  • 🐰 Snow Bunny Project: The internal model Snow Bunny has strong application-building capabilities, capable of producing 3000 lines of code in one go.

  • 🧠 Deep Reasoning Architecture: Introduces "System2" thinking and a Deep Think toggle switch, significantly improving the solution rate for complex logic problems.

  • 🚀 Multi-model Collaboration: Leaked information mentions two specialized models: Fierce Falcon (speed and logic) and Ghost Falcon (multimodal creation).

  • 📊 Performance Surpasses Competitors: In leaked benchmark tests, its reasoning score (80%) significantly outperforms competitors (55%).