Gemini 3.5 Is Google's Greatest Model Ever! Most Powerful AI EVER! (Tested)

Google’s “Snow Bunny” – A Deep Dive into the Next-Generation LLM

Key Concepts:

• Gemini 3/3.5/Pro: Google’s latest large language model family, with increasing capabilities.
• Snow Bunny: Internal codename for a new, highly capable LLM being tested by Google, potentially Gemini 3.5.
• Frontier LLMs: The most advanced large language models currently available (e.g., GPT-5, Gemini 3 Pro, Opus).
• Heroglyph Benchmark: A benchmark for evaluating lateral reasoning capabilities in LLMs.
• SVG: Scalable Vector Graphics – an XML-based vector image format used for creating icons and UI elements.
• Post Talk: A product analytics platform offering observability tools, including LLM analytics.
• Context Window: The amount of text an LLM can process at once, impacting its ability to understand and generate complex outputs.
• Aentic Coding: A coding style that prioritizes clarity and maintainability.

1. Introduction of a New Internal Model – “Snow Bunny”

Google is currently internally testing a new large language model, referred to as “Snow Bunny,” which demonstrates significant advancements over existing models like Gemini 3 Pro and even Opus 4.5. This model is speculated to be either the generally available (GA) version of Gemini 3 or potentially Gemini 3.5, with a potential release within the next few weeks. The initial demonstrations showcase an unprecedented ability to generate complex, fully functional operating systems and user interfaces.

2. Operating System Generation Capabilities

The most striking demonstration involves the generation of complete operating systems. While Gemini 3 Pro was already capable of creating a functional browser-based OS with applications like Safari, “Snow Bunny” surpasses this by flawlessly replicating a macOS operating system. This includes:

• Visual Fidelity: The generated OS is visually almost indistinguishable from a real macOS, with meticulously crafted SVG-based icons, window behaviors, and UI components.
• Functionality: The Safari application within the generated OS is fully functional, allowing users to browse the web.
• Precision: The model demonstrates an “insane level of precision” in handling structure, styling, and interactivity.
• Windows Clone: Snow Bunny also successfully generated a fully functional Windows clone, complete with functional apps and accurate icons.

3. Benchmarking and Performance Data

Developer Leo (@leo on Twitter) benchmarked “Snow Bunny” using his “Heroglyph” benchmark, which assesses lateral reasoning capabilities. The results indicate that “Snow Bunny” (both “Raw” and “Lesser Raw” variants) outperformed Lithium Flow and even Gemini 3 Pro preview in this specific task, demonstrating a significant performance gap.

Further testing involved a challenge to create a standalone HTML file rendering a realistic candle flame animation that melts over 10 seconds.

• Gemini 3 Pro: Performed well, with a functional animation and flame extinguishing.
• Snow Bunny: Improved upon Gemini 3 Pro’s animation, showcasing realistic wax accumulation.
• Kumi K2.5: Surprisingly, Kumi K2.5 generated the most impressive result, with the most realistic lighting and animation.

4. Front-End Development and AI-Generated UI

“Snow Bunny” excels at generating front-end code. A demo showcased the model creating a complete front-end layout from a descriptive prompt. The speaker emphasizes the importance of detailed prompting to achieve optimal results.

• Chhattislaw’s Demo: Chhattislaw created an Eiffel Tower in Boxil, with options for day/night, golden hour, and sparkle effects (though the sparkle effect was not fully functional).
• AI’s Increasing UI Capabilities: The speaker predicts that it will become increasingly difficult to distinguish between AI-generated and human-created user interfaces.

5. Game Boy Color Emulation

Perhaps the most remarkable demonstration is “Snow Bunny’s” ability to generate a fully functional Game Boy Color emulator in a single shot. This includes:

• Multiple Games: The emulator comes pre-loaded with games like Galactic Boy and Snake.
• Pokemon Compatibility: The emulator is potentially capable of running Pokemon games.
• Illustrative of AI’s Potential: This showcases the model’s ability to create complex, interactive systems from scratch.

6. Post Talk Sponsorship – Observability for AI

The video includes a sponsored segment for Post Talk, a product analytics platform. Post Talk provides observability tools specifically designed for AI applications, including:

• LLM Analytics: Tracing prompts, inspecting responses, and understanding AI feature behavior.
• Traditional Observability: Error tracking, session replay, feature flags, and experiments.
• Data Warehouse: Integrated CDP and data warehouse for querying product data with SQL or no-code insights.

7. Context Handling and Future Improvements

Leaks suggest potential improvements to “Snow Bunny’s” context handling capabilities, which could lead to even higher quality outputs. However, increased context windows typically come with increased computational costs. The speaker notes the potential for increased expense for both consumers and Google.

8. The Blurring Line Between AI and Human Creation

The speaker concludes by emphasizing the rapidly advancing capabilities of AI, particularly in coding and UI generation. They predict a future where it will be increasingly difficult to differentiate between AI-generated and human-created content.

Notable Quotes:

• “This part is just insanely wild because the functionality of the Safari app actually works where you're able to search through different sorts of pages.”
• “This is what really makes you stop and think when you ask yourself how much powerful an AI model can actually get. The answer is clearly a lot more.”
• “Now we're going to get to a point where we're not going to be able to tell if certain frontends are actually generated by AI or not.”

Conclusion:

“Snow Bunny” represents a significant leap forward in large language model capabilities, particularly in coding and UI generation. Its ability to create fully functional operating systems and emulators demonstrates the immense potential of AI to automate complex tasks and blur the lines between human and machine creation. The upcoming release of Gemini 3 or 3.5, potentially incorporating these advancements, promises to be a groundbreaking moment in the field of artificial intelligence.