Gemma Gemini, a free chat that shows which model answered and why

Gemma Gemini (opens in a new tab) is a free chat that looks like Google's Gemini app and needs no account. You type a message and one of two models answers it: Gemma 4 E2B, running on our NVIDIA Jetson Orin Nano boards, or Gemini 4 Argon, reached in the cloud through OpenRouter. A small router reads each message first and picks.

Gemma Gemini is an Offline Base spinoff. We built it on October 3, 2026, and we run it independently of Google, which makes both of the models it talks to.

Two models behind one text box

Gemma 4 E2B (opens in a new tab) has about 2.3 billion effective parameters and an Apache 2.0 license. We run ggml-org's Q4_0 GGUF (opens in a new tab) of it, a 2.84 GB file, under Ollama (opens in a new tab). On a Jetson Orin Nano it writes about 25 to 29 tokens per second. Inside the Jetson Orin Nano Super, the computer behind Base Router covers the hardware.

The same model is the default everyday model on Base Drive, our $19.99 USB drive, which is coming soon. Our site lists it at 3 GB on disk and 8 GB of laptop memory, and How much memory a local AI model needs explains what that means for your machine. Gemma Gemini lets you get a feel for that model now, though its speed on a laptop will differ from a Jetson's. Base Drive, private AI for the laptop you already own covers the drive itself.

For harder messages, Gemma Gemini calls Gemini 4 Argon (opens in a new tab) through OpenRouter.

How the router decides

Since October 13, 2026, EmbeddingGemma 2 (opens in a new tab), Google's open embedding model, has made the call. It turns a piece of text into a list of 768 numbers that captures what the text is about. Its model card lists 740 million parameters in total, 270 million of them in the text model. That is small enough to run on a Jetson's CPU.

Two small classifiers that we trained read those 768 numbers:

• One gives the chance that Gemma 4 E2B answers the message well. At 50% or more, Gemma answers. Below 50%, Gemini 4 Argon does.
• The other gives the chance that the message needs current information from the web. At 50% or more, the app runs a web search before the model answers.

Before shipping, we trained the classifiers on 118 labeled messages and tested them on 46 others written in a different style. They picked the right model for 93% of those. The version now running was then retrained on all 164 messages, so read 93% as an estimate from that test.

Forty-six messages is a small test set. At that size each message is worth about two points: 93% means three misses, and one more or one fewer would move the score to 91% or 96%. A larger test set will give a firmer figure.

Before October 13, the router was Jev, TypeSafe's decision model. Jev was also the router behind tryingopen (opens in a new tab), one of our free chats with open-weight models, until October 12. Our Jetson fleet has its own router, Liquid AI's d1-3B, and How a decision model picks between our Jetsons and the cloud explains how that one works.

You can also skip the router. The model menu in the text box starts on Auto, and you can pick either model yourself. If the router fails to answer in time, built-in rules decide. The trace shows both cases.

Every reply shows its route

Each reply has a chip under it that names the model and where it ran. Click it and the route trace opens:

• a small diagram of the path: you, then the router, then Gemma or Gemini
• one plain sentence on why the message went where it did
• a "Chance Gemma answers this well" bar with the 50% mark drawn on it, and a "Needs the web" bar
• who decided, how long the decision took, time to the first word, tokens per second, token counts and the model id

Here is a real one. On October 14 we sent "hi there". EmbeddingGemma 2 gave Gemma 4 E2B an 80% chance of answering it well and a 16% chance that it needed the web, so the message stayed on a Jetson. The first word of Gemma's thinking arrived 1.4 seconds after we sent it, routing included. Gemma then wrote 272 tokens at about 29 tokens per second, nearly all of them thinking before its eight-word reply.

The trace put that decision at 688 ms, measured from the app. The embedding step on the board takes about 36 ms. Most of the rest is network: the request travels from the app's server through the fleet's gateway to a board and back, and the app may first check that a Gemma board is up.

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.

Our about page (opens in a new tab) has a "How we work" section headed "Honest before impressive." It promises "No telemetry, no quiet updates, no hidden cloud calls."

Gemma Gemini makes cloud calls on purpose, so the trace keeps them in the open. Every reply that went to Gemini 4 Argon in the cloud says so, along with the numbers that sent it there.

Where your messages go

Both routes leave your own device.

• Every message first goes to the Gemma Gemini server, which we run, and requests to the Jetsons pass through the fleet's gateway.
• In Auto mode, the app sends your message to one of our Jetsons, where EmbeddingGemma 2 reads it to make the routing decision.
• A Gemma answer comes from our Jetson boards.
• A Gemini answer goes through OpenRouter to Gemini 4 Argon in Google's cloud.
• When the router says a message needs the web, the app also sends that message as a search query to a search provider (Exa, as of October 2026), and the trace shows the search.

Treat Gemma Gemini as a public demo, and type what you would be comfortable sending to an online service. For private work, the model belongs on a computer you own. Why we make local the default for private AI lays out that case, and Base Drive is the product we are building for it.

Try it at gemmagemini.com

Open gemmagemini.com (opens in a new tab), leave the menu on Auto, and send a mix of quick questions and hard ones. Then click the chip under each reply and check whether the router agreed with you. If it sent something to the wrong model, the trace shows you the numbers behind the mistake.

For the free chats on offlinebase.com itself, see Three things you can try on offlinebase.com without an account.