Offline Base: local AI for everyone

Offline Base is a small, early-stage company, and the headline on offlinebase.com says what we work on: making local AI easy for all. Local AI means the model runs on a computer you own. What you type stays on that machine, and the model keeps answering with the Wi-Fi off.

Our About page opens with "We think you should own your AI." AI has become part of daily life, so we think it should live on hardware you own, under rules you set. The page ends on the problem we want to fix: "Local AI is still too niche and nerdy." Gemma Gemini, the app this blog belongs to, is a spinoff of that work.

Five beliefs

We first published these in our newsletter, and the About page lists them today:

1. Local, open, small models are good enough. The About page cites a Stanford and Together AI study in which local models answered 88.7% of about a million single-turn chat and reasoning queries, when each query went to whichever local model got it right (more on that study). The share a local model could serve rose from 23.2% in 2023 to 71.3% in 2025.
2. Running models locally answers the data-center problem. A model on your own laptop sends no work to a data center, so it adds nothing to the grid strain and rising power bills that data centers bring. The About page also raises water and noise, and notes that the research on them is not fully established.
3. Teaching everyone data privacy is a lost cause (sort of). People will keep pasting sensitive things into the nearest chatbot. The fix we propose is a better default: local first, with heavier work sent to more capable models only when the owner approves it.
4. No one gets left behind in an AI-native world. We are young, and many of our friends want nothing to do with AI. We want them to benefit from it without feeling they are making a moral concession.
5. Open-source models are more than good enough. Frontier models are tools to call on when a question needs one. In the About page's words, "Routing is the name of the game."

We hold the site to a rule we call "Honest before impressive." Every price, spec and date on it is a current target, and the product pages say plainly that a small local model trails the largest cloud systems on the hardest questions.

Products in development

Base Drive and Base Router are in development and not yet available to order.

Base Drive

Base Drive will be our first product: a USB drive that sets up private, offline AI on the laptop you already have, for Apple Silicon Macs and Windows 11. One installer copies the models to the laptop, verifies them and opens a chat app in your browser. After setup you can unplug the drive and turn off the Wi-Fi, and the AI keeps working.

Its installer is open source, every file on the drive is listed, and setup checks published SHA-256 checksums before it copies anything. There is no account or telemetry, and the chat app talks to the model only over localhost.

Base Drive is targeted at $19.99 as a one-time purchase.

Base Router

Base Router puts the same private AI on an always-on box for the whole home. It is designed as a finished appliance with the software and AI already loaded. The plan builds it around an NVIDIA Jetson Orin Nano Super, a small computer that draws 7 to 25 watts. Our fleet runs on the same family of boards.

A third product, Decision Tree, puts a private decision model inside a small bonsai. It is in development too.

Running today

Free chats in the browser

tryingopen is our free browser chat for open models: "Try any open model. No account, nothing saved. Every chat can be shared as a link." On October 14 it listed 24 chat models. The default pick runs Liquid AI's LFM2.5-VL 3B on our Jetsons.

Since October 12, Liquid AI's d1-3B decision model has run on one of our boards. It decides whether each message stays on a Jetson or goes to a cloud model, which is GLM 5.3 Flash unless the visitor picks another. Every reply says which one answered. tryingopen: try any open model, no account has the full tour.

tryingopen grew out of the open models page on offlinebase.com, which still offers 10 open-weight models routed only to providers with zero data retention. On offlinebase.com/try, Liquid AI's LFM2.5 1.2B answers live from a Jetson Orin Nano Super, and those chats are not stored.

The Jetson fleet

Our own rack holds five NVIDIA Jetson Orin Nano boards with 8 GB of memory each. A control plane watches them and runs a gateway that speaks the OpenAI API, including chat, embeddings and a routing call. Each app gets its own API key, and the gateway sends each request to a board that has the model loaded.

In October the boards serve Gemma 4 E2B, Liquid AI's LFM2.5-VL 3B and EmbeddingGemma 2, plus "combo" models that pair a local model with a cloud one. The fleet page shows live numbers straight from the boards. At 21:39 UTC on October 14, its feed showed four of the five boards online, drawing 12.1 watts together, with 6,448 requests and about 4.17 million tokens served since September 30. The feed counted zero cloud tokens over that period.

Free keys with a daily allowance are available on request through the fleet page. Inside the Offline Base Jetson fleet goes further.

The Offline Base Letter

The Offline Base Letter is our newsletter on Substack. The newsletter page describes it as "clear notes on private AI, open models, and owning the tools you rely on," with new issues "when there is something worth saying." The first issue, "Intro to Offline Base" (August 30, 2026), introduced the team and the five beliefs.

Where Gemma Gemini fits

Gemma Gemini puts the first and fifth beliefs in front of anyone with a browser, free and with no account. Each message goes to Gemma 4 E2B on our Jetson Orin Nano boards or to Gemini 4 Argon through OpenRouter. Gemma answers when the router gives it a 50% or better chance of answering well. A chip under each reply shows how sure the router was, and opening it shows that chance as a bar with the 50% mark. On a Jetson, Gemma writes about 25 to 29 tokens per second.

Gemma's turns go through the fleet gateway like any other app's. Since October 13 the app's router, EmbeddingGemma 2, runs on one of the boards too, a day after tryingopen moved its own routing onto a Jetson. Both models' system prompts call the app an Offline Base spinoff, so either one can tell you where it comes from.

From tryingopen to Gemma Gemini traces how the app grew out of tryingopen, and Why Offline Base built Gemma Gemini explains our reasons.