Our 3B model's answers, set beside any cloud model you pick

Since October 1, 2026, when you pick a cloud model on tryingopen, your first two questions in a chat also go to LFM2.5-VL 3B, the 3B model from Liquid AI that runs on Offline Base's edge devices. A panel shows its answer beside your pick's.

LFM2.5-VL 3B is also the default pick, where d1, Liquid AI's decision model, reads each message on one of our devices and sends the harder ones to a cloud model (How the d1 router picks between our small model and the cloud explains the rule).

Our privacy policy describes the side-by-side in one sentence: "When you pick a cloud model instead, your first two questions in a chat are also answered by our devices, so you can compare the two."

What happens when you ask

When you send the first or second question of a chat with a cloud model picked, the page sends two requests at the same moment. One goes to your pick, as always. The other goes straight to one of our edge devices, where LFM2.5-VL 3B answers the same question.

That second request skips the steps a message to the default pick goes through. It makes no routing call, because the goal is to see the small model's answer whatever d1 would have decided. It runs no web search, and the panel says so: "Here it doesn't search the web, so on recent events it answers from what it already knows." For your second question, the small model reads the chat so far, which means your first question and your pick's first reply.

The small model reads pictures, so a question with a photo attached gets a second answer too.

Compare is a separate feature: it runs only when you press it under a finished reply and asks a closed model (Claude Sonnet 5 or GPT-5.6 Terra) the same question (Ask a closed model the same question with Compare).

Where the answers appear

On a wide screen

A tab labelled "Small model" waits on the right edge of the window. It flashes, because a quiet tab at the edge of the screen is easy to miss. Click or tap it, or pull it in from the edge, and a panel titled "How the small model would have answered" slides open. The page steps aside to make room, so your pick's answer stays in the chat on the left and the small model's answer sits beside it.

On a phone

A phone has no margin beside the chat, so a tab there would cover the reply. Phones get a "See how the small model answered" button under the reply it belongs to. It opens the same panel over the chat, and a swipe to the right closes it.

What the panel shows

At the top, the panel explains itself: "The first two questions in a chat also go to Offline Base's small model: 3B parameters, running on a palm-sized edge device instead of in a data center, so you can compare its answers with your pick's." Our default model runs on a palm-sized Jetson Orin Nano covers the hardware.

Then, for each question, it shows your question, the small model's full answer (credited to Liquid AI LFM2.5-VL 3B), a drawing of the two models' sizes and their timings.

Parameter counts drawn to scale

Two squares stand for the two models, and each square's area matches its model's parameter count, so a hundredfold gap looks a hundred times bigger. Beside GLM 5.3 Flash, which the picker lists at 320B, the 3B square is a speck, and the caption reads "GLM 5.3 Flash has about 107× the parameters."

The caption adds a fair warning: "Parameter count is a rough guide to size: many large models use only a slice of theirs for each word they write." GLM 5.3 Flash uses 18B of its 320B per token, and What total and active parameters tell you about a model explains why that matters. A few models carry no size in the picker (Qwen3.8 Flash, Mistral Large 4 and MiMo V2.6 Pro), and for those the panel leaves the drawing out.

Timings

Each side gets two numbers, "First words" and "Finished". The small model's are "timed from this page": the clock starts in your browser when the request leaves, so its times include the trip to the device and back. Your pick's are "timed at the server", from the same figures as the reply's "How this reply was produced" panel (What each reply tells you about how it was made).

Limits on the small model's answers

The side-by-side borrows the devices that answer default-pick chats, so its limits are tighter than a default-pick reply's.

The panel's answer is there for comparison, and while a device writes it, that device can't take anyone else's message. That's why the cap sits well below the default pick's 4,000. If you use up the allowance, the panel says how many side-by-sides the small model has answered for you this hour and when the count resets.

Chats that skip the side-by-side

Self Chat. In Self Chat, a message goes either straight to OpenRouter under a zero-data-retention requirement or to a model on your own computer (plus one web lookup, only if you switch search on). Our devices sit on a different path, through an ngrok tunnel, so Self Chat never sends them a copy.

The two edge-device picks. When you pick the default LFM2.5-VL 3B or Gemma 4 E2B, d1 already decides for each message whether a device answers, and every reply says which one did. The side-by-side runs only when your pick is a cloud model.

What the devices keep

Our privacy policy says the devices record the size and timing of requests and keep no message text. Your first two questions make one extra trip, through the ngrok tunnel to our devices, and the privacy policy sentence quoted near the top of this post discloses it. What tryingopen keeps, and what stays on your device covers the rest.

Why we show a 3B model beside your pick

Pick GLM 5.3 Flash and the panel puts a 3B model on a palm-sized device next to a 320B model in a data center. Reading both answers to your own questions shows which everyday questions a 3B model already handles.

On the default pick, d1 makes that call on every message. It keeps greetings, short facts, short rewrites and pictures on the small model, and sends code, multi-step maths and longer, detailed answers to the cloud. The side-by-side lets you check that call on questions d1 never scored.