Google Gemini leaks show hidden models

Gemini Live looks like it’s getting split into multiple hidden personalities just before Google I/O. That matters because Live is the part of Gemini people actually talk to — by voice, with camera and screen sharing — and right now it can feel like one broad model trying to do every job. The new leak suggests Google is carving that into narrower modes with different strengths. If that holds, I/O on May 19-20 may be less about one giant “Gemini 4” reveal and more about packaging several behaviors behind a cleaner consumer interface. (forbes.com) ### What actually leaked? The leak came from a hidden model selector inside Google App v17.18.22. Forbes says the menu exposes seven Gemini Live model options that are not normally visible to users, including a new “Thinking” variant and several names that do not appear in earlier public documentation. The key point is simple — these are not just strings buried in code. The reporter says he enabled them and got measurably different behavior from different options. (forbes.com) ### Why is Gemini Live the interesting part? Gemini Live is Google’s real-time, conversational layer — the thing meant to feel less like a chatbot prompt box and more like an always-on assistant. Google already pitches Live as voice-first and multimodal, with camera and screen sharing in the Gemini app. So if Google is testing hidden Live-specific models, that points to a product decision, not just a lab experiment. Basically, Google seems to be tuning the assistant for different kinds of live interactions instead of forcing one model to handle everything. (apkmirror.com) ### What are the seven models? The most concrete details are the count and the broad differences, not a full official list from Google. Forbes says the hidden selector shows seven options, including a “Thinking” model, two RC2 variants, and three codenames that appear to be new in this context — “Capybara,” “Nitrogen,” and a personalization variant. That mix matters because it hints at three tracks at once: reasoning, pre-release candidate testing, and user-tailored behavior. (forbes.com) ### Are they really different, or just labels? Turns out there are functional differences. Forbes says four of the hidden models could access the user’s location and return live weather, while three could not. That sounds small, but it’s a big tell. If some models can pull live context and others can’t, Google is not merely renaming the same engine. It’s testing different permission scopes, tool access, or product roles — like one model for grounded assistant tasks and another for slower reasoning. (forbes.com) ### Why would Google hide this now? Because I/O is nine days away, and Google has already said the conference will feature Gemini model updates and product demos. Hidden selectors in beta or production apps often show up right before launches because teams need to test combinations without exposing unfinished features to everyone. That doesn’t guarantee all seven models ship. But it does suggest Google is in late-stage integration, where the demo story and the backend model choices are starting to line up. (blog.google) ### Does this mean Gemini 4 is coming? Maybe, but the leak itself does not prove that. A lot of pre-I/O coverage is framing the event around a possible Gemini 4 announcement, but Google’s own I/O materials only promise “latest Gemini model updates,” not a named flagship launch. The safer read is that Google is preparing a more segmented Gemini product stack, and a new top-line model could be one part of that rather than the whole story. (developers.googleblog.com) ### What’s the real takeaway? The interesting shift is not just “seven hidden models.” It’s that Gemini may be moving from one assistant with one brain to one assistant with a router behind it. Think less one super-model, more a switchboard — one mode for quick live replies, one for deeper reasoning, one for personalized context, maybe one for tool use. That is how these products get faster and more useful without making every interaction expensive or slow. (forbes.com) ### So what should people watch at I/O? Watch for anything that changes how Gemini Live presents itself — new mode names, a visible “thinking” toggle, more proactive personalization, or demos where Gemini clearly switches between fast answers and slower agent-like tasks. If Google shows that kind of handoff cleanly, this leak will look less like a curiosity and more like an early map of the product strategy. (forbes.com) The bottom line is that this leak points to a more modular Gemini, not just a bigger one. And for users, that is probably the more important upgrade.