Introducing Gemma Gemini
A free chat app where a router sends each message to Gemma 4 E2B on a Jetson Orin Nano or to Gemini 4 Argon in the cloud, and shows you why.
Gemma Gemini puts two of Google's models behind one text box. Gemma 4 E2B, a small open-weight model, runs on NVIDIA Jetson Orin Nano boards that we own. Gemini 4 Argon, Google's new frontier model, runs in Google's cloud. A router reads each message and picks which one answers.
The app looks like Google's Gemini app, down to the sidebar, the model picker, the light and dark themes and the Google Sans Flex typeface, and it works on phones. We built the first version on October 3, 2026.
Why two models in one chat
Much of what people type into a chatbot is everyday: a greeting, a definition, a translation, a quick fact. We think a small open model answers most of these well. Gemma 4 E2B on a Jetson writes about 25 to 29 tokens per second, and with the model warm its first word usually arrives in about 0.2 to 1.3 seconds.
Other messages need a much bigger model, like a proof or a question about mixing medications. Google's release notes call Argon "our most capable model built for complex coding, agentic, math, and science tasks." It is slower. In our testing, a full proof with reasoning effort set to high took about 14 seconds end to end.
What you see
The picker in the composer, titled "Who answers", has three settings. Auto is the default and lets the router choose. Picking Gemma 4 E2B or Gemini 4 Argon skips the router and sends your messages to that model, except that images and PDFs always go to Gemini. A redo menu on the latest reply can let the router pick again or send the same message to the other model. The new answer replaces the old one.
Each reply ends with a chip naming the model and where it ran, such as "Gemma 4 E2B · Jetson Orin Nano", followed by the router's verdict, such as "EmbeddingGemma 80%". The percentage is how sure the router was of the side it picked. Click the chip to open the route trace:
- a diagram: you, then the router, then Gemma or Gemini
- a "Chance Gemma answers this well" bar with a mark at 50%, and a "Needs the web" bar
- a one-sentence explanation in plain English
- decision time, time to first word, tokens per second, token counts and the model id
Gemma answers when its chance is 50% or higher. Some scores from the live router:
| Message | Chance Gemma answers this well | Answered by |
|---|---|---|
| "hi there" | 80% | Gemma 4 E2B |
| "What's the weather in NYC today?" | 94% | Gemma 4 E2B, after a web search |
| "Prove that the square root of 2 is irrational" | 20% | Gemini 4 Argon |
| A question about taking ibuprofen with blood pressure medication | 23% | Gemini 4 Argon |
Search, thinking and failover
When the router gives a message a 50% or higher chance of needing the web, the app runs an Exa search first and hands the results to whichever model answers. That is how a small model on a Jetson can tell you today's weather.
Both models show their reasoning in a collapsible "Show thinking" section. If the chosen model fails before its first word, for example because a board is offline, the other model answers and the trace says so. Images and PDFs go to Gemini, because the Gemma setup on our boards reads text only.
Where the models run
Gemma 4 E2B is the smallest of Google's open-weight Gemma 4 models, with 2.3 billion effective parameters. We run ggml-org's 4-bit Q4_0 GGUF of it, a 2.84 GB file, under Ollama on NVIDIA Jetson Orin Nano boards in Offline Base's fleet. Each board has 8 GB of memory shared by its CPU and GPU. The fleet has five boards, and as of October 14, two of them serve Gemma.
Since October 13, the router has run on a Jetson too. EmbeddingGemma 2, Google's open embedding model, turns each message into 768 numbers on one of our boards, and two small classifiers in the app turn them into the two chances in the trace.
It replaced Jev, TypeSafe's decision model, which we reached through a cloud API. Jev scored a little higher on our 46 held-out test messages (98% against 93%). We switched anyway, so the model reading your message runs on hardware we own. If the router fails or takes longer than 2.5 seconds, built-in keyword rules decide.
We reach Gemini 4 Argon through OpenRouter as google/gemini-4-argon, with reasoning effort set to high. Google announced Argon on September 30 and made it generally available on October 13, the day we moved to it from Gemini 3.8 Flash.
Free, with chats saved in your browser
Gemma Gemini is free and public, with no account and no password. Chats are saved only in your browser's local storage, so a chat link opened in another browser says "This chat isn't in this browser." Settings has a button to delete them all.
To answer you, the app's server sends Gemma turns to our Jetson fleet and Gemini turns through OpenRouter to Google. In Auto mode the router embeds each message on one of our boards, and a turn that needs the web sends your latest message to Exa.
Where Gemma Gemini comes from
Gemma Gemini is a spinoff of Offline Base, an early-stage company that builds local AI: models that run on hardware you own, so what you type stays on your machine and keeps working offline.
Offline Base is developing Base Drive, a USB drive that sets up private, offline AI on the laptop you already own (targeted at $19.99, paid once), and Base Router, an always-on box that brings the same private AI to the whole home. Both are in development and not yet available to order.
Offline Base's About page puts it this way: "Routing is the name of the game." Gemma Gemini puts that idea in front of anyone with a browser.
Offline Base also runs tryingopen, where you can try open models with no account and nothing saved. Gemma Gemini's first router was modeled on the one in tryingopen, and both apps now run their router models on Offline Base hardware: Liquid AI's d1-3B for tryingopen, EmbeddingGemma 2 for Gemma Gemini.
Read more
Gemma Gemini sends each message to Gemma 4 E2B or to Gemini 4 Argon, and shows you why.
Try Gemma Gemini