Choosing an engine

A new chat asks you to pick a model to run: the chip beside Send reads Pick a model to run until you do. The picker lists every model on the shelf; the ones your account can't use right now are greyed, with the reason. Vresk remembers your last pick in this browser, so the next new chat starts from it when it is available.

Worth weighing when you pick:

• The work is demanding. A designed deck or a substantial build asks much more of a model than a question does.
• You care where the model was built. See origins below.
• You already know what you like. If you have found an engine that suits how you write, pin it.

One capability is worth knowing before you pin: engines that can read images carry a mark on the shelf and answer a search for “photo” or “vision”, and pinning one that can't on a conversation that has images gets a refusal that names the problem — never a quiet swap to a different engine. More on images.

Choose the engine at the start of a piece of work, not sentence by sentence. A conversation is usually one task, and one task usually wants one engine — a new engine mid-thought spends its first move re-reading everything above it, which costs you time and buys very little.

What actually needs to travel between engines is not conversation state. It is the layer that belongs to you: what Vresk remembers about how you work. That layer is deliberately kept separate from any one thread so it is there whichever engine you start with. More on memory.

Every completed answer carries a small engine badge. Tap it and a popover tells you which engine answered and where it was built — read from the answer's own record, not reconstructed after the fact. When the turn needed more than a straight answer — a research run, a web search, a designed document, a picture, tools — it says that too.

It also says whether you picked that engine for this conversation. That line is read from the conversation itself, never guessed; if Vresk cannot check, the popover says it cannot tell you.

From the same popover, one tap keeps answering with the engine that just impressed you. A pick applies from the next message on.

On a phone the badge is the model's name under each answer. Under the latest answer that name opens What I used instead (what that reply was given to work from), so on a phone this popover, and its pin, open only from an earlier answer. In a chat with one answer there is no earlier one, and on a phone neither can be reached until the second answer arrives; the model chip still switches models there.

Every model carries a badge for where it was built — United States, Europe, China. It is there so the choice is yours to make on information rather than on assumption.

The default shelf is global. The Model regions toggle on the picker — All, or North America and Europe — narrows it to models built in North America and Europe. Engines outside that line stay on the shelf but stop being selectable: they fold behind a single row that says how many are hidden, and opening it shows each one, greyed, with a note saying why. You see what you are excluding rather than wondering what the shelf is hiding. Flipping it either way is entirely your call.

Wherever a model was built, your conversations are not used to train it. More on data.

The curated set of engines, with the plain-English note the picker shows for each. Which of these are selectable for you at a given moment depends on your origin setting and on what is switched on at the time — the picker is the live answer.