Run several models at once

One question goes to several models at the same moment. The answers come back side by side. Big-AGI calls the run a Beam; the optional step that combines the answers into one reply is a merge.

Why Beam

For high-stakes answers, one model is not enough. One answer gives you nothing to check it against: whether it hallucinated, whether you can bet your business on it. In a multi-step chat, taking the right turn at the right time matters more than polishing any single reply. Several models address the question from different angles, and that reveals blind spots. Where they agree, you gain confidence in the answer. Where they disagree, you know exactly where to dig (reading disagreement). Beam turns uncertainty into confidence, and approximation into quality.

How it works

Beam opens its own view over the chat. Your message sits at the top with everything attached to it - documents, images, the conversation so far. Under it, one card per model: model, persona and parameters are set per card. Start them all with one press, or each card individually. When the answers are in, remove the ones you don't want, then merge: four kinds of merge, the fourth your own custom operation. Accept any answer, accept the merge, or discard everything. The one message that goes back to the chat is the one you picked.

Prompt out, one card per model, an optional merge, one message back.

Against four browser tabs: one prompt, one history, one screen, and a merge over every answer.

What a run costs

What you run	Requests
4 models on the panel	4 requests, one per card, each on that model's AI service
each merge card	1 more request; Guided is 2
an imported answer	0 - copied from the chat, not regenerated
a run you stopped	what was already written is already billed

Every card issues its own complete chat request. The merge re-sends the conversation plus every answer, the largest request of the run. Every request goes on your AI service's bill; Big-AGI adds no fee.

When it pays off

Early and short. Every turn re-sends the history, so the multiplier costs most deep in a long thread. The in-app tutorial says best used in earlier / shorter chats. When being wrong is expensive: an architecture, a diagnosis, a contract clause. One model is enough for mechanical work - reformatting, extraction, a draft you will rewrite anyway.

Beam is the same in every version, Open included, on your own keys: all four merge modes, no server component.

Related

• Combining answers
• When the models disagree
• Connect an AI service