Forbes finds seven Gemini Live models

Google’s voice assistant story just got more interesting. A teardown of the Android Google app suggests Gemini Live is no longer being treated as one model with one personality. Instead, Google seems to be testing a small lineup of hidden voice models with different strengths — and that matters because voice assistants only feel useful when they respond fast, remember context, and don’t make obvious mistakes. The gap, basically, is that one model rarely does all of that well at once. What changed this week is that Forbes surfaced a hidden selector inside Google App v17.18.22 showing seven Gemini Live variants that appear to behave differently. ### What was actually found? The report describes a hidden settings button in the Google app that exposes a model selector for Gemini Live. Inside it were seven options, including names like “Capybara,” “Nitrogen,” RC2 variants, a personalization model, and a “Thinking” path. The selector is reportedly behind a server-side flag, which usually means Google has the plumbing in place but has not turned it on for normal users. (forbes.com) ### Why does seven models matter? Because Gemini Live today is usually understood as one voice experience sitting on top of one main live model. Google’s own developer docs frame Live as a low-latency audio model for real-time dialogue, with current public model listings centered on Flash Live variants built for speed. If Google is testing seven internal options, that suggests the company is tuning for tradeoffs instead of pretending one setup fits every conversation. (forbes.com) ### What differences showed up? The interesting part is that these weren’t just different names. Forbes says the models produced measurably different responses in testing. Four could access location and return live weather. Three could not. One variant, “Capybara,” reportedly identified itself as “Gemini 3.1 Pro” instead of the standard Flash Live family. Another path leaned into personalization and memory. That is the tell — Google seems to be mixing speed, reasoning, tool access, and memory in different combinations. (ai.google.dev) ### Why would Google split voice AI this way? Because voice is a brutal product problem. A fast answer feels smart even when it is shallow, but a thoughtful answer often takes too long for spoken conversation. Then add memory, live search, app context, and emotional tone, and the system gets even harder to balance. Google’s Live API already supports things like tool use, proactive audio, multilingual switching, and affective dialog, so the platform is built for specialized behaviors. (forbes.com) The hidden models look like product packaging catching up with the underlying tech. ### Is this connected to earlier Gemini Live clues? Yes — pretty directly. Back in January, another teardown showed Google preparing “Thinking Mode” and “Experimental Features” for Gemini Live, with hints about multimodal memory, better noise handling, camera-aware responses, and personalized results from Google apps. A March report then pointed to “Personal Intelligence” coming to Live, using past conversations and connected apps like Gmail, Photos, Search, and YouTube. The new seven-model selector looks like the next layer of that same shift. (docs.cloud.google.com) ### Does this mean users will get a model picker? Maybe, but that part is still inference. The selector exists internally, and Forbes argues the infrastructure for switchable voice models is already there. But APK teardowns are snapshots of work in progress, not launch promises. Google could expose these as user-facing choices, or it could keep the picker hidden and route people automatically based on task — like using one model for fast chat and another for deeper reasoning. (9to5google.com) ### Why does this matter beyond Google? Because it hints at where voice assistants are going next. The old dream was one universal assistant. The newer reality is a stack of specialists behind one interface. That is cleaner for the model makers, and honestly better for users — if the handoff is invisible. ### Bottom line? This leak matters less because there are seven models and more because Google seems ready to admit that live AI needs modes. Fast, thoughtful, personalized, tool-using, maybe camera-aware — those are different jobs. (forbes.com) Gemini Live looks like it is being rebuilt around that fact.