Our default model runs on a palm-sized Jetson Orin Nano

Tryingopen's default pick, LFM2.5-VL 3B, answers quick questions on hardware we own: as of October 2026, a fleet of five NVIDIA Jetson Orin Nano boards.

Each board fits in one hand. The small model is a 3B vision-language model from Liquid AI. Both are described on the site itself: under any reply from the default pick, a Learn more link opens a dialog called "How the small model answers". Its second tab, "The edge device", is where most of the quotes below come from.

An 8 GB NVIDIA Jetson Orin Nano

The dialog's "The edge device" tab says:

The edge device is an NVIDIA Jetson Orin Nano: a palm-sized computer with 8 GB of memory shared by its CPU and a 1,024-core GPU, drawing 7 to 25 watts. Offline Base runs five of them, and whichever board is free takes your message.

NVIDIA brought Super mode to the Orin Nano developer kit as a software update, and its Jetson Orin Nano Super Developer Kit page lists the 7 to 25 W range.

When every board is busy, the turn goes to the cloud model. If the fleet hasn't answered within 4 seconds, the cloud model also takes the turn, "so the page keeps working with every edge device unplugged."

Liquid AI's LFM2.5-VL 3B

The small model is Liquid AI's LFM2.5-VL 3B, published on August 12, 2026. The dialog says it is "stored at 4-bit precision so it sits in the board's memory with room to spare, and served by Ollama on the board itself."

Liquid says the model runs in "about 3 GB of memory", which leaves about 5 GB of the board's 8 GB for the rest of the system.

Liquid's model card is direct about where it fits. It recommends the model for "single-turn, high-throughput, low-latency tasks" and advises against long-context, reasoning-intensive ones. We built the default pick around that advice. Short questions, pictures and lookups after a web search stay on the board. Code, maths, advice and long answers go to a cloud model you choose (GLM 5.3 Flash unless you pick another).

The router that makes that call is d1, Liquid AI's small decision model. Since October 12, 2026, it has run on one of the boards, so Offline Base's own hardware decides between board and cloud. How the d1 router picks between our small model and the cloud covers what it reads and how it decides.

You can run the same model at home (its 4-bit build is a 1.67 GB file) and talk to it from tryingopen's Self Chat. Use Ollama, LM Studio, llama.cpp or Jan from Self Chat gives the command and the setup.

How a message reaches a board

The dialog's section "How your message gets there" reads:

From this page to Offline Base's control plane through an encrypted tunnel, then down a connection the board's own agent keeps open. The boards accept no incoming connections, and the model listens only on the board.

The privacy policy names the tunnel: it runs through ngrok.

Limits on the board

The small model works inside a few fixed limits:

• 4,000 tokens per reply. The site caps edge replies there "so one answer can't hold a board for long."
• Text replies only. The board gives the model no tools, so every edge reply is plain text. Some cloud models in the picker can create files.
• Web results come from the site. When d1 says a turn needs the web, the site runs one search and puts the results into the small model's prompt. How web search works on tryingopen has the details.
• PDFs go to the cloud. Only the cloud path reads PDFs, so a turn with a PDF attached skips the board.

What the fleet keeps

The dialog's "What is kept" section says the fleet "logs each request's model, token counts and timing, never the text of your message or the reply." The privacy policy gives the same rule.

For an edge turn, which travels through the ngrok tunnel to our own devices, that logging rule is the promise. The default pick as a whole sits outside the site's zero-retention promise, because a message crosses that tunnel to reach d1 and the board, and a cloud turn or a web search goes to outside companies.

With no provider picked, GLM 5.3 Flash goes to Z.ai's own API, which may keep what you send. Pick a provider beside the cloud-fallback picker and the cloud turns go through OpenRouter with zero retention instead. Choosing the company that runs your model explains how. What zero data retention covers, and where it stops has a table of every kind of turn and what each one keeps.

Default since September 30

The edge pick became tryingopen's default on September 30, 2026. Before that, new chats started on GLM 5.3 Flash, which now answers the harder turns as the cloud fallback. What each reply tells you about how it was made walks through the labels that say which model answered.

A second small model on the same boards

The picker has a second edge entry on the same boards: Google's Gemma 4 E2B, text only, with Gemma 4 31B as its starting cloud fallback. Gemma 4 on tryingopen, from our boards to the cloud covers how it runs and when it hands off.

If you pick a cloud model instead, the boards still show up: your first two questions in a chat also go to the small model, and you can open its answers beside your pick's. How the side-by-side compares our small model with your pick shows what you see.

To see all of this for one of your own messages, ask the default pick a short question and open Learn more under the reply.