Google’s Next-Gen AI Leaked? A First Look at Gemini 3.0 Flash
A potential leak from Google has surfaced, giving us a tantalizing glimpse into what might be the next major evolution in the Gemini series. Two new mysterious models, codenamed Skyhawk and Seahawk, have appeared on the AI testing platform LM Arena, and they are showing remarkable capabilities. Speculation is high that we are seeing an early version of Gemini 3.0 Flash, a model poised to be a cost-efficient, lightning-fast powerhouse.
[0:04.269]
This exciting development comes from a new leak hinting at Gemini 3.0 Flash, a model designed to be a highly efficient and speedy variant within the Gemini 3.0 model series. We’ve only recently been introduced to the flagship Gemini 3.0 Pro, which already set a new standard with its incredible coding abilities, advanced reasoning, and versatile performance. The appearance of these new test models suggests Google is moving at an accelerated pace.
[0:16.899]
The recent release of the flagship Gemini 3.0 Pro already impressed the AI community with its state-of-the-art performance. It demonstrated elite coding capabilities, next-level reasoning skills, and unbelievable versatility across a wide range of tasks. This model alone represents a significant leap forward for Google’s AI development.
[0:21.059]
Now, it appears Google is already in the testing phase for its next iteration. New checkpoints are being evaluated within LM Arena, a public platform where AI models can be tested and compared head-to-head. These new models could be the anticipated Gemini 3.0 Flash or potentially even an early variant of a future Gemini 3.5, indicating that Google is using these platforms to discreetly benchmark its upcoming technology.
[0:34.909]
[0:37.389]
[0:43.349]
Currently, two new models are being tested within LM Arena. This platform allows anyone to autonomously test and compare different AI models in real-time battles. Tech companies often use this to gauge their new models against existing competitors like those from Mistral AI. In one test, a prompt was given to create a YouTube clone. One of the new models, Seahawk, generated a high-quality, functional front-end for a YouTube clone with impressive speed and accuracy.
[1:12.199]
The generation quality from the Seahawk model is genuinely powerful. It produced a comprehensive YouTube clone in a single pass, featuring a responsive layout, video grid with thumbnails, search functionality, and a sleek dark mode theme using Tailwind CSS. The ability to generate such a complete project so quickly is a testament to the model’s advanced capabilities and a very promising sign of what’s to come.
[1:38.979]
Users can test these new models, Skyhawk and Seahawk, through LM Arena. There are also reports that some users are gaining access to these new checkpoints through Google AI Studio and the main Gemini web app via random A/B testing. While access through these official channels is rare, it confirms that Google is actively rolling out these tests to a wider, albeit select, audience.
[2:05.109]
This isn’t just wild speculation. Sundar Pichai, the CEO of Google, has publicly commented on the upcoming model, adding weight to these leaks.
Gemini 3.0 Flash is coming and it’s gonna be a very very good model. Might be our best one yet. ~ Sundar Pichai
This statement suggests that Gemini 3.0 Flash is not just a scaled-down or “light” version. Google seems to be positioning it as a genuine breakthrough—a model that achieves top-tier intelligence while being significantly faster and cheaper to run, thanks to its specialized “flashed architecture.” It’s designed to be a game-changer, not a compromise.
[2:51.109]
To try and access these new variants, you can head over to LM Arena and select the “Battle” mode. This pits two anonymous models against each other. By using the “Code” mode and providing a complex prompt, you might get lucky and have one of your options generated by Skyhawk or Seahawk. After you vote on which output is better, the platform reveals the names of the models used.
[3:43.319]
In one such test, the Skyhawk variant was prompted to create a browser-based OS, and the result was extraordinary. It generated a sophisticated operating system interface complete with a desktop environment, clickable icons, a file explorer, notifications, and several functioning applications, including a System Monitor, a Web Browser, and a Terminal. The quality and complexity of this output from a single prompt are deeply impressive.
[4:17.659]
The codenames Skyhawk and Seahawk are consistent with Google’s internal naming conventions, further strengthening the belief that these are indeed Google’s new models. The code generated is clean, extensive, and leverages modern frameworks like Tailwind CSS. The model’s ability to handle such a large and complex request, generating multiple interconnected files and functionalities, showcases a powerful understanding of long-context instructions.
[4:44.279]
To push the models further, another prompt requested a “polished, creative weather app with vivid, 3D-animated weather types taken to the extreme.” The Skyhawk model produced a visually stunning and interactive weather application called “AETHER.EXT.” It features beautiful, animated simulations for various extreme weather scenarios like “Solar” flares, “Cyclones,” and “Acid” rain, changing dynamically based on the selected city. The UI/UX design is professional grade, demonstrating a high level of creativity and technical execution.
[5:48.719]
In a final, highly complex test, the models were asked to build a “realistic, working, and accurate 3D mechanical watch mechanism.” Here, the Seahawk model truly shined. It created a detailed, interactive 3D simulation of a watch’s inner workings. Users can explore different layers and components like the Balance Wheel, Escapement, and Mainspring Barrel. Hovering over each part reveals a description of its function. While a competing model from Anthropic (claude-sonnet-4.5) failed to place the gears correctly, Seahawk delivered a functional and educational visualization, showcasing its superior spatial and logical reasoning for complex mechanical systems.
[6:55.779]
With this level of performance seen in testing, the official release seems imminent. Prediction markets like Polymarket show traders betting heavily on a December 2025 release for Gemini 3.0 Flash. Given the advanced state of these leaked models, it’s highly likely that Google will officially launch this powerful, efficient, and surprisingly capable AI before the year is out, setting a new benchmark for what we can expect from AI models.