Google tests seven Gemini Live

- Google appears to be testing seven hidden Gemini Live variants inside the Android app, with different behaviors for speed, reasoning, personalization, and voice flow. - The clearest tell is the hidden selector itself: app probes surfaced names like Capybara, Nitrogen, Thinking, Personalization, and two RC2 builds. - That matters because Gemini Live already runs on 3.1 Flash Live, so extra variants suggest Google is splitting live tasks across specialized models.

Google’s Gemini app looks like it’s growing a routing layer — not just a bigger base model. The new leak is about Gemini Live, the real-time voice mode, and the interesting part is that Google may be testing seven hidden variants behind the scenes. If that holds up, Gemini Live stops being “one voice AI” and starts looking more like a traffic controller that sends different parts of a conversation to different systems. That would be a big shift for how live assistants actually work. ### What leaked this time? A hidden model selector showed up in Google app version 17.18.22, and reporting tied it to Gemini Live rather than the regular text chatbot. The list included seven options, with names and labels suggesting distinct modes rather than duplicate builds. Some of the names surfaced publicly were Capybara, Nitrogen, a personalization variant, a Thinking model, and two RC2 entries. The catch is that this came from teardown work and internal testing paths, so none of it guarantees a public launch. (forbes.com) ### Why would Gemini Live need seven versions? Live voice is a messy job. One part has to hear you fast. Another part has to decide what you mean. Another may need to look at the camera feed or screen share. Another may need to remember your preferences without slowing everything down. A single giant model can do all of that, but usually with tradeoffs — more latency, higher cost, or worse turn-taking. Multiple specialized models are the cleaner engineering answer. That’s why the leak matters. (forbes.com) It hints Google is optimizing for the task, not just the benchmark. ### What do the names suggest? The names are clues, not documentation. “Thinking” sounds like a slower, more deliberate mode for harder questions. “Personalization” points to answers shaped by your Google context or past preferences. The RC2 labels look like release-candidate testing. Capybara and Nitrogen are harder to decode, but the reporting says the variants behaved differently in use, which matters more than the names. Basically, Google seems to be testing whether users notice better live behavior when the system swaps models under the hood. (9to5google.com) ### Didn’t Google already upgrade Gemini Live? Yes — and that’s part of why this is interesting. In March, Google said Gemini Live was getting its “biggest upgrade yet” with Gemini 3.1 Flash Live, pitched as its highest-quality audio and voice model so far. That upgrade already covered faster voice interaction and live audio-video back-and-forth in Gemini and Search. So these seven hidden variants do not look like a first draft. They look more like the next layer on top — finer-grained routing after the main model upgrade landed. (forbes.com) ### How does the video leak fit in? A separate leak points to a Gemini “Omni” video model, with early demos showing prompt-based video generation and editing inside Gemini. That appears to be a different product path from Live voice, but the timing is revealing. Google seems to be pushing Gemini toward one interface that can talk, see, edit, and generate across formats. Live variants would handle real-time interaction; Omni-style video tools would handle heavier media creation jobs that do not need instant replies. (9to5google.com) ### Is this definitely coming at I/O? Not definitely. But the calendar lines up. Google I/O 2026 is set for May 19-20, and these leaks landed just days before the event. That is exactly when unfinished app strings and hidden toggles tend to surface. The safest read is that Google is actively testing the architecture now, with some part of it likely to be previewed soon — even if the exact seven-model setup never ships as-is. (chromeunboxed.com) ### What’s the real takeaway? The important change is not the number seven. It’s the idea that Gemini Live may be turning into a system of cooperating models instead of one assistant voice with one brain. If Google pulls that off, the user experience gets simpler while the stack underneath gets much more complicated. That’s usually the sign a product is getting serious. (chromeunboxed.com)

Get your own daily briefing

Scout delivers personalized news, insights, and conversations tailored to your role and industry.

Download on the App Store

Shared from Scout - Be the smartest in the room.