跪拜 Guibai
← All articles
GitHub

Leaked Gemini 3.5 Pro Shows One-Shot Frontend Generation That Rivals Fable 5

By Maynor在掘金 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A model that generates production-quality frontend code from a single prompt changes the speed at which UI prototypes and marketing pages get built. The retrained base also signals Google is willing to delay a flagship launch to swap foundations entirely, raising the floor for what a delayed model must deliver.

Summary

Google's Gemini 3.5 Pro, rumored for a July 17 debut after a two-month delay, has surfaced in leaks showing a sharp leap in frontend and visual code generation. Demos circulating on X depict one-shot complete UIs, complex hand-drawn SVG orreries, and Three.js steampunk islands — outputs with professional design taste, clean structure, and fewer redundant elements than typical AI-generated pages. Developers are calling the visual dominance 'mogging' over rivals.

The delay stems from a full retraining on a new base, not a simple scale-up of Gemini 3.5 Flash. Leaks also point to a companion image model, Nano Banana Pro, built on the same foundation to compete with OpenAI's GPT-Image 2. Google is using one retrained base to feed two product lines simultaneously.

Despite the frontend flash, leaked benchmarks place Gemini 3.5 Pro behind Fable 5 and GPT-5.6 on repository-level engineering, multi-step reasoning, and complex agent tasks. It's a specialized, sharp tool for visual output, not an all-around leader.

Takeaways
Gemini 3.5 Pro was delayed because Google retrained it from scratch on a new base, rather than scaling up the Flash variant.
Leaked demos show one-shot generation of complete UIs, complex SVGs, and Three.js scenes with professional-level design taste.
On visual and frontend tasks, the leaked Pro version outperforms Fable 5 in side-by-side comparisons.
Hard reasoning, repository-level engineering, and long-horizon agent tasks remain weaker than Fable 5 and GPT-5.6.
Google is also preparing an image model called Nano Banana Pro on the same retrained base, targeting GPT-Image 2.
The rumored release date is July 17, though Google has not confirmed it officially.
Conclusions

Google's willingness to scrap an existing base and retrain from scratch for a point-release suggests internal benchmarks showed the old foundation was a dead end for the gains they wanted.

Specialization is becoming the norm: Gemini 3.5 Pro doesn't try to win everywhere, but aims to dominate a specific, high-visibility capability — frontend generation — where demos sell the product.

The companion image model launch on the same base reveals a platform strategy: one expensive retraining event amortized across multiple product lines to close gaps with OpenAI on two fronts at once.

Concepts & terms
Mogging
Slang originating from looksmaxxing communities, meaning to dominate or overshadow someone in appearance or capability. Here it describes Gemini 3.5 Pro's visual output clearly outclassing competitors.
SWE-Bench Pro
A benchmark that evaluates AI models on real-world software engineering tasks, including repository-level debugging and complex code modifications, used to measure agentic coding ability.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗